AI & Machine Learning

AlloyDB AI Functions - now with revolutionary performance boosts and cost savings

Wed, 01 Jul 2026 18:00:00 +0000

AlloyDB is an AI-native database—it isn’t just a passive data store, it intelligently understands and processes your data. With AlloyDB, you get industry-leading vector and hybrid search, near 100% accurate natural language-to-SQL capabilities to build conversational agents, tools to enable you to build with your agentic IDEs of choice, and the ability to bring the intelligence of foundation models like Gemini directly to your data through AI functions.

In this blog post, we discuss the massive breakthroughs in AI function processing alongside a suite of brand-new AI functions.

But first: what exactly are AI functions? They bring Gemini’s world knowledge to your AlloyDB data. Consider the challenge of managing raw user feedback: it’s unstructured, and difficult to parse through. Before this data can be leveraged for search, it may require pre-processing and entity extraction. Rather than maintaining complex custom pipelines for knowledge extraction, you can use Gemini’s generation capabilities directly within AlloyDB to transform raw text into structured, searchable insights. For example, here is how you can use ai.generate to instantly turn raw feedback into clean, structured JSON (see more examples here):

code_block: <ListValue: [StructValue([('code', "SELECT\r\n log_id,\r\n raw_content,\r\n -- Use Gemini 3.0 to reason through the raw user feedback and extract structure\r\n ai.generate(\r\n model_id => 'gemini-3.1-pro-preview',\r\n prompt =>\r\n 'Analyze this raw customer feedback entry. Extract the country, service name, and a 1-sentence summary of the feedback. Return as JSON.'\r\n || raw_content) AS structured_feedback\r\nFROM raw_feedback_logs\r\nWHERE user_type <> 'internal';"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa655f432e0>)])]>

Here is a sample result:

log_id	raw_content	structured_analysis
1001	2025-12-16 08:00:01 [ERROR] Service: OrderSvc \| DbConnectionTimeout: Failed to acquire connection from pool "primary-shard-04" after 5000ms.	{"errorCode": "DbConnectionTimeout", "serviceName": "OrderSvc", "rootCause": "The service failed to acquire a database connection from the primary shard pool within the 5000ms timeout limit."}
1002	2025-12-16 08:05:12 [WARN] Service: IdentityProvider \| 401 Unauthorized: Bearer token validation failed for user_id=9942. Signature mismatch.	{ "error_code": "401", "service_name": "IdentityProvider", "root_cause": "The bearer token validation failed due to a signature mismatch." }
1003	2025-12-16 08:12:45 [CRITICAL] Service: AnalyticsEngine \| OutOfMemoryError: Java heap space. Allocation of 1.2GB array failed. Heap usage 99%.	{ "error_code": "OutOfMemoryError", "service_name": "AnalyticsEngine", "root_cause": "The service exhausted available Java heap memory attempting to allocate a 1.2GB array." }
1004	2025-12-16 08:25:33 [ERROR] Service: WebFrontEnd \| 404 NotFound: Resource /api/v3/users/profile/settings not found. Upstream returned 404.	{ "error_code": "404", "service_name": "WebFrontEnd", "root_cause": "The requested API resource for user profile settings was not found by the upstream service." }
1005	2025-12-16 08:35:50 [WARN] Service: NotificationGateway \| GatewayTimeout: External provider "SendGrid" failed to respond within 30s. Retry scheduled.	{"error_code": "GatewayTimeout", "service_name": "NotificationGateway", "root_cause": "The external provider SendGrid failed to respond within the 30-second timeout limit."}

More functions to summarize and analyze sentiment

Our core AI functions —ai.generate, ai.rank, ai.if, and ai.forecast—are now Generally Available. To learn more about use cases for the first three, refer to this blog post. To explore the forecast function in action, check out this deep dive.

Building on this momentum, we have introduced three brand new functions: ai.summarize, ai.agg_summarize, and ai.analyze_sentiment.

ai.analyze_sentiment: Automatically classifies the emotional tone of text as positive, negative, or neutral.
ai.summarize: Condenses lengthy text into its most essential information while preserving the original tone and nuance.
ai.agg_summarize: An aggregate tool that processes multiple rows within a column to generate a single, unified summary for an entire group (e.g., via a GROUP BY clause).

Here’s an example of how to use ai.agg_summarize to consolidate a product reviews for products on a retail website:

code_block: <ListValue: [StructValue([('code', 'SELECT productname, ai.agg_summarize(review) as reviews_summary\r\nGROUP BY productname;'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa655f43100>)])]>

Here is a sample result of summarized reviews for two gaming console products:

productname

reviews_summary

AlphaCore Console

Users praise the stunning 4K graphics, smooth 120Hz frame rates, and the highly ergonomic controller design.

However, several reviews express frustration over the loud cooling fan noise during extended gaming sessions.

Overall, it is considered a top-tier console despite minor thermal and noise complaints.

NeoCore Console

Customers love the exceptional battery life and vibrant OLED display for handheld gaming on the go.

A significant number of users noted that the UI can feel sluggish and the game library is currently limited.

It represents great value for casual gamers but power users may find the performance lacking.

The power of LLMs on your data: now significantly faster and cheaper

We now have achieved unprecedented performance and cost breakthroughs in AI function processing. Previously, running a foundation model call for every single row in a massive database introduced cost and latency constraints. We have shattered these barriers by introducing two breakthrough capabilities:

Smart Batching for AI Functions: This AI Function Acceleration capability provides intelligent batching of AI function calls for optimal performance and quality. This efficiency is achieved by deduplicating prompt overhead; the LLM's boilerplate instructions are transmitted once per batch rather than repeated across every individual row. A question you may have is - “Why not do this in my own application layer?”. That’s because, AlloyDB intelligently determines the right batch size for optimal results - if you underestimate the batch size, you won’t reap gains for cost and latency, and if you overestimate the batch size, the prompt to the LLM could get bloated and lead to hallucinations, or you could exceed the model's token limits. In addition to calculating the perfect batch size for every request, AlloyDB also handles retries automatically out of the box, ensuring your pipeline stays resilient. We did some testing internally and saw massive gains; for example, an up to 2,400x performance boost (processing 10,000 rows/sec) over traditional row-at-a-time LLM calls. This is currently available for the ai.if and ai.rank functions, with support for additional functions coming in the future.

Let’s look at an example of using Smart Batching / Acceleration with ai.if to solve this use case: Imagine a customer on a gadget retail site searching for a camera that can handle an underwater depth of '60 meters or deeper.' Traditional hybrid search will pull the closest semantic and full-text matches, but it misses the hard constraints of numerical data—meaning it might serve up a camera that works only at 20 meters depth. By using AlloyDB’s ai.if-based intelligent filtering, the database actually understands the nuance of depth and makes the query return products that meet or exceed that 60-meter depth criteria. Notice how, in the example below, you don’t need to specify the batch size - AlloyDB handles all the optimizations under the hood when using ai.if.

code_block: <ListValue: [StructValue([('code', "-- Smart Batching / AI Function Acceleration \r\nSET google_ml_integration.enable_ai_function_acceleration = on;\r\nSELECT productid, productname, category,description\r\nFROM products AS p\r\nWHERE\r\n ai.if(\r\n 'Evaluate if the product description indicates that the product is waterproof at depth 60m or deeper. Description:'\r\n || description);"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa655f43760>)])]>

Here is a sample result on a hypothetical gadgets site. Notice how the expanded descriptions of products really match the criteria of working at a depth of 60 meters:

Optimized AI Functions: For even greater efficiency, we’ve introduced an optimized mode, starting with ai.if. By deploying a small, proxy model that utilizes your embeddings and is trained on your specific LLM outputs, we can process decisions natively within the database. This drastically reduces the need to call the external LLM - and based on some of our internal tests, we saw staggering gains; for example, up to 100,000 rows processed per second (a 23,000x improvement) and costs slashed by 6,000x (down to 1/10th of a cent). For technical insights on this technique, including when it works best and when not, refer to this blog post. AlloyDB does the following when using optimized ai.if:

Trains a proxy model: AlloyDB trains a lightweight proxy model on a sample of your data. This happens in the background when you use the PREPARE statement with ai.if function to train the model for optimized queries.
Executes the query: When you use the EXECUTE statement, AlloyDB uses the trained proxy model to process the query locally.
Falls back to the LLM: If the accuracy of the model is low, or if AlloyDB can't find a model, AlloyDB automatically falls back to using the LLM.

Let’s look at the same example of searching for a camera that can handle an underwater depth of 60 meters or deeper using optimized ai.if. Here we train a proxy model using the PREPARE statement and then EXECUTE the statement thereafter.

code_block: <ListValue: [StructValue([('code', "-- Prepare the Optimized Function / Proxy Model\r\nPREPARE waterproof_camera_60m AS\r\nSELECT productid, productname, category, description\r\nFROM products AS p\r\nWHERE\r\n ai.if(\r\n 'Evaluate if the product description indicates that the product is waterproof at depth 60m or deeper. Description:'\r\n || description,\r\n description_embedding);\r\n\r\n-- Run the Proxy Model\r\nEXECUTE waterproof_camera_60m;"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa655f43160>)])]>

You see the same products that truly match the criteria of working at a depth of 60 meters - as shown in the screenshot above. Here’s a tabulated version for the first three products, so you can look at the descriptions more closely:

productname	description
Pulsetron Action Camera MZ314	Conquer your next adventure with this camera. Don't let the elements hold you back; dive up to 60 meters deep or withstand rugged trails with its shock-resistant, adventure-ready chassis. Every jump, every turn, every splash is rendered flawlessly smooth with advanced Horizon Lock stabilization, ensuring your footage tells the story with unparalleled fluidity.
Hyperbyte Action Camera LG688	Capture the world in breathtaking detail, even when the action is at its most intense. This camera packs a formidable 1-inch sensor into a remarkably tough, pocket-sized frame. Shoot stunning 5K video and crystal-clear 20MP stills that rival professional equipment. Dive deeper than ever before with robust waterproofing at 60 meters.
Alphasync Action Camera WW897	This formidable, compact camera shrugs off the elements, while the massive 1-inch sensor translates every breathtaking moment into stunning 5K video and crystal-clear 20MP stills. Conquer any environment – from the deepest dive to the highest peak – thanks to its 60 meter waterproofing and revolutionary Horizon Lock, ensuring your footage remains impossibly steady.

See it in action!

Watch how this all comes together in this demo video.

Getting started is easy

Ready to bring unprecedented speed and cost-efficiency to your AI workloads?

New to AlloyDB? Discover AlloyDB with a 30-day free trial.
AI functions quickstart: Enable a few quick prerequisites and start calling functions like ai.if, ai.generate, or ai.analyze_sentiment directly within your SQL queries. Check out these practical examples to begin.
Boost performance and optimize costs: To unlock the biggest performance and cost gains, follow our guide on optimized functions. This is available in preview for ai.if, and will be expanding to more functions soon. For technical insights on this technique, including when it works best and when not, refer to this blog post.
Scale your throughput: Use smart batching to accelerate AI functions (available in preview for ai.if and ai.rank) or array-based functions (generally available for all LLM-based AI functions) to handle bulk prompting smoothly.

Get started with the Claude apps gateway for Google Cloud

Wed, 01 Jul 2026 16:00:00 +0000

Anthropic's agentic coding tool Claude Code has worked with Google Cloud for a while now. An individual developer could easily point CLAUDE_CODE_USE_VERTEX=1 at a Google Cloud (GCP) project, grant the role roles/aiplatform.user, and inference stays inside your Google Cloud perimeter.

That flow works great when it’s just you, or a handful of engineers. But rolling it out across an organization forces you to deal with enterprise friction: you have to manage per-developer cloud credentials, push a managed-settings.json to every laptop over MDM, and not be verified with zero per-developer usage attribution or easily enforceable spend caps.

The Claude apps gateway closes that gap. It is a self-hosted service, shipped with the same claude binary, that sits directly between your local Claude Code clients and Google Cloud. This post breaks down exactly why you should run it and what a secure deployment looks like on Google Cloud.

(Note: If you want to jump straight to the code, the full walkthrough lives in the Claude apps gateway on Google Cloud docs.)

Why run the gateway

Run the gateway to centralize the governance that developers and platform admins otherwise each carry alone such as identity, policy, cost, and routing. Here's what that looks like in practice.

Identity. The /login request routes through your identity provider (IdP ) - Google Workspace or any OIDC/OpenID Connect one - and the gateway swaps the token for a short-lived session. No sensitive information lands on the developer’s laptop — such as service-account keys, API keys, or ANTHROPIC_VERTEX_PROJECT_ID. Onboarding is as simple as adding a user to an IdP group; offboarding by removing them, and their next session refresh fails on the spot.

Policy. Your RBAC (role-based access control) rules live once in gateway.yaml, resolved per group and enforced server-side. The gateway re-checks availableModels on every /v1/messages call, so editing local managed-settings.json changes nothing — and rule updates reach the whole fleet within the hour.

Telemetry. Every claude_code.token.usage metric carries the verified email and groups from the session JWT (signed session token), not the spoofable client-set OTEL_RESOURCE_ATTRIBUTES. The gateway ships them over OTLP/HTTP to a collector you run — Cloud Monitoring, Grafana, Datadog, whatever you use.

Spend limits. Set daily, weekly, or monthly caps per user, group, or org via the admin API; the gateway meters tokens against a Cloud SQL ledger and returns a 429 at the cap. Costs are at list price, so treat them as a runaway-usage guardrail, not a bill reconciliation (committed-use discounts and negotiated rates don't show up).

Routing. Calls go out under a single Cloud Run service identity. Set region: global for Agent Platform's global endpoint, or add a second upstreams: entry to fail over on 5xx/429/timeout in list order. Either way, inference stays in your GCP project — quota, Data Processing Agreement, and billing all unchanged.

How it fits together

A developer's local or deployed claude process sends inference traffic to the gateway over HTTPS. The gateway is a stateless container on Cloud Run as shown below.

The gateway validates its own session bearer — Google Workspace is only contacted at sign-in and token refresh — checks policy, and forwards the request to Agent Platform using the Cloud Run service account. Cloud SQL holds device-code sign-in state and the spend ledger; an OTLP collector receives the attributed metrics.

Setting it up on Google Cloud

The full walkthrough, every gcloud command and the complete gateway.yaml reference, is in the Claude apps gateway on Google Cloud docs. The short version:

Step 1: Provision the GCP foundation
Enable the Agent Platform, Cloud SQL, and Secret Manager APIs; create a claude-gateway service account with roles/aiplatform.user; stand up a small Cloud SQL Postgres database instance for state. The gateway authenticates to Agent Platform as the Cloud Run service identity — you do not create a service-account key. Finally, create a new OAuth client (type Web application) in the Google Cloud console: in this example, the gateway authenticates developers against Google Workspace as an OIDC relying party, and this client is what issues it a client_id and client_secret for that handshake. Those two values feed the oidc: block in the next step. You'll later add the authorized redirect URI once the gateway URL is known.

Step 2: Configure the gateway
Write gateway.yaml pointing at your Google Workspace OIDC client, the Postgres connection string, and Agent Platform as the upstream. Store it in Secret Manager, along with the OIDC client secret, the Postgres URL, and a JWT signing key.

code_block: <ListValue: [StructValue([('code', 'listen:\r\n port: 8080\r\n public_url: https://<your-cloud-run-service-url> # the Cloud Run service URL — with --ingress=internal this resolves only inside your VPC / corporate network\r\noidc:\r\n issuer: https://accounts.google.com # Google Workspace\r\n client_id: <client-id>.apps.googleusercontent.com\r\n client_secret: ${OIDC_CLIENT_SECRET} # from Secret Manager\r\n allowed_email_domains: [yourco.com]\r\n\r\nupstreams:\r\n - provider: vertex\r\n region: us-east5\r\n project_id: <your-project>\r\n auth: {} # ADC via the Cloud Run SA, NO key file'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa655ba4970>)])]>

Then register https://<public_url host>/oauth/callback as an authorized redirect URI on the Google OAuth client — it must match listen.public_url exactly:

Step 3: Deploy to Cloud Run
gcloud run deploy with the service account attached, the Cloud SQL connection on the VPC, and the config mounted from Secret Manager. The container is stateless and scales horizontally behind the Cloud Run load balancer. GKE works equally well if that's already your platform, and only the deployment manifest changes.

code_block: <ListValue: [StructValue([('code', 'gcloud run deploy claude-gateway \\\r\n --service-account="claude-gateway@${PROJECT_ID}.iam.gserviceaccount.com" \\\r\n --set-secrets=/etc/claude/gateway.yaml=gateway-config:latest \\\r\n --ingress=internal \\ # private — developers reach the gateway over the corporate network (VPN/Interconnect into the VPC)\r\n --no-invoker-iam-check # the gateway runs its OWN OIDC; clients carry no GCP token'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa655ba4ee0>)])]>

Developers connect over the corporate network; you may front the service with an internal Application Load Balancer — see Cloud Run private networking.

Either public or internal, your developers must be able to access whatever URL you configure or you can rely on the default URL from Cloud Run. For the below example we will use https://claude-gateway.example.internal

Step 4: Onboard a developer
Push forceLoginMethod: "gateway" and forceLoginGatewayUrl to developer machines via managed settings. This is how /login knows where to connect, with no manual URL entry. For an org rollout, that's your MDM channel. For a first trial without MDM, the developer can write the file by hand at /Library/Application Support/ClaudeCode/managed-settings.json on macOS (or /etc/claude-code/managed-settings.json on Linux) if they have local admin permissions:

code_block: <ListValue: [StructValue([('code', '{\r\n "forceLoginMethod": "gateway",\r\n "forceLoginGatewayUrl": "https://claude-gateway.example.internal"\r\n}'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa655ba4eb0>)])]>

At Claude Code startup, the developer then presses Enter on the pre-filled gateway sign-in screen to confirm the URL.Confirm the device code on the gateway's verification page in the browser, and get redirected to Google Workspace to sign in. After that, the developer completes the device-code flow in the browser against Google Workspace. If setup ends correctly, you will be able to see Cloud Gateway in the terminal view as shown below.

What's next

At this point you should have a better understanding of how to configure and use Claude apps gateway on Google Cloud. Here are some next steps you may want to consider:

Full config reference: every gateway.yaml field is in claude-apps-gateway-config. Per-IdP setup and the GKE track live in claude-apps-gateway-deploy and claude-apps-gateway-on-gcp.
Group-scoped policies: front the gateway with a groups-capable IdP, set groups_claim, and add match: { groups: [...] } policies above the catch-all to give different teams different model lists and tool permissions.

For now, thanks for reading! And if you have any additional questions or feedback, feel free to reach out on socials (Roy Arsan - Linkedin, X and Ivan Nardini - LinkedIn, X)

Happy building!

What Google Cloud announced in AI this month

Tue, 30 Jun 2026 16:00:00 +0000

Editor’s note: Want to keep up with the latest from Google Cloud? Check back here for a monthly recap of our latest updates, announcements, resources, events, learning opportunities, and more.

Our main focus in June was helping your teams build, scale, and secure AI. Today, we’re sharing a fresh roundup of updates designed to help you run smarter, more secure applications while keeping everything under your control.

We even shared a cool virtual shopping demo at Cannes to show how retailers can make product discovery more exciting. Let’s dive in!

Top announcements

Introducing the Open Knowledge Format: We introduced the Open Knowledge Format (OKF), an open specification that formalizes the LLM-wiki pattern into a portable, interoperable format. This is a vendor-neutral, agent- and human-friendly standard for representing the metadata, context, and curated knowledge that modern AI systems need.
Collaboration with Apple on its expanded Private Cloud Compute (PCC) systems: Our collaboration with Apple is built on a foundation of deep commitment to privacy that leverages Google Cloud's security and privacy technologies. At the heart of this collaboration is our Confidential Computing portfolio and our Titanium security architecture.
Claude Fable 5: Available on Google Cloud: Claude Fable 5, Anthropic’s latest frontier model, is now generally available on Google Cloud. This launch is the latest proof point of our ongoing commitment to bring the industry's latest models straight to our Agent Platform.
Cloud Atelier: How Gemini Enterprise is helping restyle the retail playbook: This year at Cannes, we showcased Cloud Atelier — a destination-based, virtual shopping experience that highlights how retail brands can turn this classic dilemma into an exciting moment of product discovery.

Thought leadership (editor’s pick):

How Google Cloud Security uses AI internally: To counter machine-speed, AI-driven threats, we’ve worked hard to transition Google Cloud’s security posture to an autonomous, proactive model. By embedding specialized AI agents directly into our software development lifecycle (SDLC), we’ve created automated guardrails that protect code at a scale and speed unreachable by human teams — and we’re taking steps to make those same guardrails widely available.
The 4 lessons that guided AI Threat Defense: We introduced Chris Betz as the new CISO of Google Cloud. For his first Cloud CISO Perspectives, Chris shares four key lessons we learned about using AI to the defender’s advantage while building AI Threat Defense.

News you can use:

5 lessons from red teaming AI applications: To help you build AI securely, Mandiant has developed a proactive, risk-based approach centered on the Good AI Assessment (GAIA) Top 10, outlined in our new report, Secure Development of Generative AI Applications: A Proactive Approach.
How to unlock true ROI in software development – a deep dive into the latest DORA research: To help you evaluate the costs and business benefits of AI, we recently shared the DORA: ROI of AI-assisted software development report. This research offers a practical approach to help your team work through early adoption challenges, align engineering plans, and drive business growth.
Agent Factory Recap: 100X engineering with AI agents in Google Antigravity 2.0: In this episode of the Agent Factory, Shir Meir Lador, Head of AI Engineering, Google Cloud Developer Relations, sat down with Rody Davis, one of Google’s top agentic engineers. They dive into the massive shift from traditional IDEs to agent-first platforms, the reality of code reviews in an AI-driven world, and how to use "skills" to perform at a 100X level.

Stay tuned for monthly updates on Google Cloud’s AI announcements, news, and best practices. For a deeper dive into the latest from Google Cloud customers, read our monthly recap, Cool stuff customers built.

aside_block: <ListValue: [StructValue([('title', '$300 in free credit to try Google Cloud AI and ML'), ('body', <wagtail.rich_text.RichText object at 0x7fa6550304f0>), ('btn_text', 'Start building for free'), ('href', 'http://console.cloud.google.com/freetrial?redirectPath=/vertex-ai/'), ('image', None)])]>

May

We’ve had a busy month! Between announcing Gemini Spark and Gemini 3.5 at Google I/O – and unveiling Google AI Threat Defense, our latest AI-powered cybersecurity solution, we had a lot to share with Google Cloud customers. Keeping up with the latest news takes time, so we gathered the most important announcements, thought leadership, and technical guides in one place to help you quickly catch up.

To learn more about our I/O announcements, here’s everything you need to know for Google Cloud customers, and top news for startups.

Top announcements

Introducing Google AI Threat Defense to help you outpace the adversary: Google Cloud is introducing a comprehensive AI-powered cybersecurity solution — Google AI Threat Defense — an always-on autonomous security platform. Learn more here.

Gemini 3.5: Our latest family of models combines frontier intelligence with action – starting with Gemini 3.5 Flash.
Gemini Omni: Our new model is a leap forward in world understanding, multimodality, and editing, letting you generate any output from any input, starting with video.
Google Antigravity: Google Antigravity’s expanded capabilities and new integration with Agent Platform bring agentic development to your entire organization.
Gemini Spark: For Gemini Enterprise and Workspace customers, Gemini Spark is your 24/7 personal AI agent that helps you work more efficiently by autonomously taking action on your behalf, under your direction.
Google Workspace: Google Pics, our new image generation and editing tool, and new voice features in Gmail, Docs and Keep, help reimagine how you work.
Managed Agents API on Agent Platform: Allows developers to build and run custom agents inside secure, Google-hosted environments that seamlessly integrate with Agent Platform.
CodeMender: A powerful AI security agent provided through Agent Platform, CodeMender can help find and fix vulnerabilities in your code.

Nano Banana 2 and Nano Banana Pro are generally available: Available today via Gemini Enterprise Agent Platform, organizations are already putting the models to work. Learn more here.

Thought leadership (editor’s pick):

Cloud CISO Perspectives: How Google + Wiz changes multicloud strategy for CISOs: Vinod D’Souza, director, Office of the CISO, shares highlights from his RSA Conference fireside chat with Anthony Belfiore, chief strategy officer, Wiz. While threat actors have seen gains from the adversarial misuse of AI, Google and Wiz are tackling these challenges head-on by combining Wiz's deep cloud telemetry with Google's world-class AI and quantum research to help CISOs and their organizations meet the needs of the agentic enterprise era. Read more here.

News you can use:

What Google I/O '26 means for developing agents on Google Cloud: Dig deep into how Gemini Enterprise Agent Platform and the new developer tools shared at I/O fit together, unpack the spectrum of choice for building, and share what we’d actually try first. Learn more here.

Five must-have guides to move agents into production with Gemini Enterprise Agent Platform: Here is a look back at our five-part series covering the architecture patterns and best practices you need to move your agents into production. Learn more here.
How to build an AI-ready security program for the public sector: From industrial control systems to decades-old municipal databases, here’s our CISO guidance to prep AI-ready security programs for the public sector. Learn more here.

April

We hosted Google Cloud Next in Las Vegas on April 22, announcing incredible innovations from Gemini Enterprise Agent Platform to our eight-generation TPUs. We also expanded the Gemini Enterprise app in collaborative ways – now, with new features like Projects, you can work side-by-side with your agents and colleagues.

If you missed the livestream, take a look at our Day 1 recap. It’s been incredible to see how customers have been applying AI in thousands of ways — so far, we’ve counted more than 1,300 examples.

Top announcements

1. Gemini Enterprise Agent Platform: Our new, comprehensive platform to build, scale, govern, and optimize agents. Moving forward, all Vertex AI services and roadmap evolutions will be delivered exclusively through the Agent Platform, rather than as a standalone service, to power the next generation of agent development.

The platform is designed around four core pillars — build, scale, govern, and optimize — that allow teams to collaborate seamlessly. Learn more about Agent Platform here.

2. Gemini Enterprise app has all the key components to let teams discover, create, share, and run AI agents in a single environment. At Next ‘26, we introduced several new capabilities in the Gemini Enterprise app:

Agent Designer uses the same no-code agent designer experience of Agent Platform and lets employees build sophisticated schedule- and trigger-based agents using any enterprise connector. It gives you a virtual flowchart of your agent, allowing you to inspect, test, and approve workflows, ensuring total transparency for executing critical business processes.
Long-running agents are designed to execute complex business processes. They can work autonomously in secure cloud sandboxes, giving agents the ability to orchestrate business logic, write code to build custom tools, and complete multi-step work like reconciliation activities or sales prospect sequencing — without needing constant prompting.
Inbox in Gemini Enterprise provides a central location to monitor, guide, and help manage all of your agent activity, including your long-running agents. Notifications are intuitively categorized into actionable groups like "Needs your input," "Errors," and "Completed.”
Projects create a dedicated space where the agent’s memory is confined to the files and conversations your team adds. By connecting it to data sources including Google Drive, NotebookLM, and Google Group Chats, the agent becomes an expert on a specific topic and can provide team members daily briefings or status updates without digging through months of documents.
Skills create simple shortcuts using an “@” mention for repetitive tasks such as applying brand guidelines, formatting a report, and accessing specific data.
Canvas gives our customers an interactive editor directly within Gemini Enterprise. It allows teams to easily create and edit Docs and Slides, and even export to Microsoft 365 files, within the same experience.
Agent Gallery provides access to third-party agents from partners like Adobe, Atlassian, Lovable, and ServiceNow, and is adding more third-party connectors for Asana, Mailchimp, Workday, and more. These integrations enable your agents to retrieve data and execute tasks with your systems-of-record.

3. AI Hypercomputer: Designed specifically for demanding AI workloads, our AI Hypercomputer is an advanced, purpose-built architecture that unites performance-optimized hardware for compute, storage, networking, open software and machine learning frameworks — as well as flexible consumption models — into a single, integrated system. We are announcing innovations at every layer of the AI Hypercomputer:

TPU 8t, optimized for training, uses breakthrough Inter-Chip Interconnect (ICI) technology to scale up to 9,600 TPUs and 2 PB of shared, high-bandwidth memory in a single superpod. It achieves 3x the processing power of Ironwood and delivers up to 2x more performance/Watt.
TPU 8i, optimized for inference, uses our new Boardfly topology to directly connect 1,152 TPUs in a single pod. It features 3x more on-chip SRAM compared to previous versions to host larger KV caches entirely on-silicon and integrates a specialized Collectives Acceleration Engine. Taken together, TPU 8i delivers 80% better performance per dollar for inference than the prior generation, enabling millions of concurrent agents to run cost-effectively.

4. The Agentic Data Cloud: A new data architecture built for the speed and scale of agentic AI. The Agentic Data Cloud delivers an AI-native architecture, allowing agents to perceive, reason, and act on your behalf in real-time, including:

Cross-Cloud Lakehouse, standardized on Apache Iceberg, is our Lakehouse that enables you to leave your data in AWS or Azure (coming later this year) while querying it instantly — without the friction of vendor lock-in or the cost of data movement
Knowledge Catalog constructs a unified, dynamic context graph of your entire business enabling you to ground agents in all of your business data and semantics. With Smart Storage and the Object Context API, files in Google Cloud Storage are instantly tagged and enriched with metadata before an agent touches them. Then our Knowledge Engine uses Gemini to autonomously tag, define logic and instantly map complex relationships across your entire enterprise, providing the semantic definition your agents have been missing.

5. Protecting the agentic enterprise: Security built for the AI era. Our full-stack AI approach, from the chips to the models, gives you a competitive advantage with better integration and velocity to help protect customers. Not only can Google action insights from the world’s largest threat observatory and Mandiant frontline experts, but we also bring cutting-edge insights and breakthroughs from Google DeepMind, to help make your platforms more secure.

Agentic defense: Three new agents in Google Security Operations can help hunt threats, engineer detections, and provide context on third parties. You can build your own security agents with remote Google Cloud model context protocol (MCP) server support for Google Security Operations, now generally available. You can also access the MCP server client directly from the Google Security Operations chat interface, available in preview.
Protecting AI and cloud apps across any infrastructure with Wiz: Newly expanded AI coverage helps build secure agents across clouds and AI studios. New AI-Bill of Materials in development tools can help secure AI-generated code and mitigate the risk of shadow AI. Learn more.
Securing agents and the agentic web: Model Armor can integrate with Agent Gateway, and new Agent Identities provide more layers of defense against shadow AI. Google Cloud Fraud Defense, the next evolution of reCAPTCHA, offers agent-specific capabilities that can help secure the agentic web as well as the entire user and customer journey.
Trusted Cloud: We’re simplifying permissions with modern IAM, and advancing Google Cloud security with new capabilities in Security Command Center plus new innovations in data and network security.
New partner-supported workflows for Google Security Operations: This new robust cohort of partner integrations includes partners developing their own agentic security operations centers (SOCs).

You can catch up on all our security announcements from Next ‘26 here.

News you can use

Guide to prompting Gemini 3.1 Flash TTS (text-to-speech): The new TTS model introduces a high level of controllability by allowing you to steer the delivery using more than 200 audio tags. We'll share how to get strong results from the model, whether you are building accessible gaming soundtracks, banking systems, or audiobooks. Learn more about the model here.
Ultimate prompting guide for Lyria 3 models: Lyria 3, Google's family of music-generation models, is designed to give you granular control over vocals, instrumentation, and arrangement. So we spent weeks testing against every musical genre and use case we could imagine. We put together this guide to share exactly what we learned and how you can get the best results.
How to find the sweet spot between cost and performance: This guide will walk you through Google Cloud's flexible gen AI infrastructure options, showing you how to find that sweet spot on the efficient frontier between cost and performance. We'll start with the foundational pay-as-you-go (PayGo) models and then explore how to layer on more specialized options to build a robust and cost-effective gen AI strategy.
Essential AI and cloud security now on by default: To support the next generation of AI innovators, we are offering on by default essential AI security and cloud security in Security Command Center Standard.
Securing AI inference on GKE with Model Armor: Here’s how to secure AI inference on Google Kubernetes Engine with Model Armor and high-performance storage.
Cloud CISO Perspectives: AI, security, and the workforce of the future: You can’t bring traditional security to an AI fight, so how do we defend against AI-powered attacks, boost defenders with AI, and secure AI use? Drop in on this RSA Conference fireside chat between Francis deSouza, Google Cloud COO and President, Security Products, and Nick Godfrey, senior director, Office of the CISO.

March

March was a busy month for our AI teams. We launched Gemini Embedding 2, rolled out a highly cost-effective Veo 3.1 Lite model, and officially welcomed the Wiz team to Google Cloud to help redefine security in the AI era.

Alongside these launches, we created comprehensive guides to help you get the most out of these models, from prompting formulas for Nano Banana 2, to practical advice for optimizing your TPU training. Here’s a quick look at the latest news and resources to help your team build what’s next.

Top hits:

Gemini Embedding 2: Our first natively multimodal embedding model: Gemini Embedding 2 is our first natively multimodal embedding model that maps text, images, video, audio and documents into a single embedding space, enabling multimodal retrieval and classification across different types of media — and it’s available now in public preview.
Build with Veo 3.1 Lite, our most cost-effective video generation model: This model empowers developers to build high-volume video applications, at less than 50% of the cost of Veo 3.1 Fast, but with the same speed. This rounds out the Veo 3.1 model family, giving developers flexibility based on needs. For Cloud customers, it’s now available on Vertex AI.

Here’s a fun bonus: Check out our ultimate prompting guide for Veo 3.1 to get started.

Welcoming Wiz to Google Cloud: Redefining security for the AI era: Google has completed its acquisition of Wiz, a leading cloud and AI security platform. The Wiz team will join Google Cloud, and we will retain the Wiz brand. With the addition of Wiz, we will provide customers with a comprehensive platform to secure their cloud and hybrid environments, as well as accelerate threat prevention, detection, and response.
Gemini 3.1 Flash Live: Making audio AI more natural and reliable: We’ve improved 3.1 Flash Live’s overall quality, making it more reliable for developers and enterprises to build voice-first agents that can complete complex tasks at scale. On ComplexFuncBench Audio, a benchmark that captures multi-step function calling with various constraints, it leads with a score of 90.8% compared to our previous model.

News you can use:

The ultimate Nano Banana prompting guide: This is a must-read for anyone working with Nano Banana. We spent weeks testing Nano Banana 2 and Nano Banana Pro against every use case we could imagine to test its limits. We put together this guide to share exactly what we learned and how you can get the best results. Here’s an example formula: [Reference images] + [Relationship instruction] + [New scenario]

A developer’s guide to training with Ironwood TPUs: In this guide, we hear from Lillian Yu, CPA, CA , Product Strategy and Operation, and Liat Berry, Product Manager, on five strategies within the JAX and MaxText ecosystems designed to help developers refine training efficiency and hit peak performance on Ironwood hardware.
How to build production-ready AI agents with Google-managed MCP servers: In this guide, we anchor on a specific example. Cityscape is a demo agent built with Google's Application Development Kit (ADK) that turns a simple text prompt — like "Generate a cityscape for Kyoto" — into a unique, AI-generated city image. Check out the guide to learn more.

February

In February, we’re giving developers more reasoning power with Gemini 3.1 Pro and Claude 4.6, and faster creative scaling with Nano Banana 2. We’re also opening up new training programs and step-by-step guides to help you tackle the hardest parts of the AI lifecycle, from capacity planning to mounting defenses against AI-powered attacks.

Here’s a rundown of our latest news, tools, and resources to help you build what’s next.

Top hits

Pro-level image generation gets faster and more accessible with Nano Banana 2: To build creative that stands out, you need models that naturally integrate into your workflows and scale with ease. Check out our blog to see how this comes to life (and how customers are putting the model to work).

Introducing Gemini 3.1 Pro on Google Cloud: Gemini 3.1 Pro is a clear step forward in reasoning, designed to solve tougher problems, giving you the reasoning depth your business needs. Gemini 3.1 Pro is available starting today in preview in Vertex AI and Gemini Enterprise. Developers can access the model in preview via the Gemini API in Google AI Studio, Android Studio, Google Antigravity, and Gemini CLI.
Announcing Claude Opus 4.6 and Claude Sonnet 4.6 on Vertex AI: Now generally available on Vertex AI, explore our sample notebook to get started and visit our documentation for comprehensive pricing and regional availability details.
New AI threats report: Distillation, experimentation, and integration: John Hultquist, chief analyst, Google Threat Intelligence Group, details what security leaders should know from our newest AI threat report on experimentation, integration, and distillation attacks.

News you can use

A developer's guide to production-ready AI agents: To help developers work through these challenges, we've published a collection of guides covering the full agent lifecycle. These resources first appeared during Kaggle’s 5 days of AI Agents Intensive, and they’ve proven so popular and useful, we wanted to make sure a wider audience had access, as well.
Gemini Enterprise Agent Ready (GEAR) program now available: We opened the Gemini Enterprise Agent Ready (GEAR) learning program to everyone. As a new specialized pathway within the Google Developer Program, GEAR empowers developers and pros to build and deploy enterprise-grade agents with Google AI.
Your guide to Provisioned Throughput (PT) on Vertex AI: Check out this deep-dive blog designed to show you the resources available to you today on Vertex AI, and how you can get started capacity planning.
How AI can boost defenders, from defense in depth to the cyber kill chain (Q&A): We know that defenders are also developing powerful AI tools, but what’s still unknown is what it could mean for enterprise software ownership if companies have to constantly mount AI-directed defenses at AI-powered attacks?

Janurary

We used to have to learn the language of computers. In 2026, they’re learning ours.

We kicked off the year by exploring the future of agentic commerce, where AI agents navigate the web to find and buy products for us. Our leaders call this the "invisible shelf" — a world where commerce isn't tied to a specific website. To make this reality scalable, we announced the Universal Commerce Protocol (UCP), a shared language that allows agents and retailers to understand each other.

We brought that same fluency to our creative and technical tools:

Updates to Veo 3.1 allow creators to use simple inputs — like reference images — to generate precise, mobile-ready video.
Natural language queries: With Comments to SQL in BigQuery, we’re removing the language barrier to data. Engineers can now write queries by describing their intent in natural language, prioritizing the question over the code.

Let’s dive in.

Top hits

1. Gemini Enterprise for Customer Experience (CX): Specifically built for agentic retail, this platform transforms fragmented search, commerce and service touch points into one seamless journey — whether you need a shopping assistant, a support bot, agentic search or help with merchandising.

2. We announced Universal Commerce Protocol (UCP): A new open standard for agentic commerce that works across the entire shopping journey — from discovery and buying to post-purchase support. UCP establishes a common language for agents and systems to operate together across consumer surfaces, businesses and payment providers. So instead of requiring unique connections for every individual agent, UCP enables all agents to interact easily. UCP is built to work across verticals and is compatible with existing industry protocols like Agent2Agent (A2A), Agent Payments Protocol (AP2) and Model Context Protocol (MCP).

3. We updated Veo 3.1, including improvements to Ingredients to Video and Portrait mode: Veo is getting more expressive, with improvements that help you create more fun, creative, high-quality videos based on ingredient images, built directly for the mobile format. This includes:

Improvements to Veo 3.1 Ingredients to Video, our capability that lets you create videos based on reference images.
Native vertical outputs for Ingredients to Video (portrait mode) to power mobile-first, short-form video creation.
State-of-the-art upscaling to 1080p and 4K resolution 1 for high-fidelity production workflows.

These updates are launching in the Gemini app, YouTube, Flow, Google Vids, the Gemini API and Vertex AI.

4. Vibe querying with comments-to-SQL: Crafting complex SQL queries can be challenging. Often, engineers simply want to express their data needs in plain English directly within their SQL workflow. That’s why we’re introducing Comments to SQL in BigQuery. This feature makes writing queries using natural language – ‘vibe querying’ – a reality. Learn more in the blog.

News you can use

Mastering Gemini CLI: Your complete guide from installation to advanced use-cases: We’ve teamed up with DeepLearning.ai and are excited to announce a free course – Gemini CLI: Code & Create with an Open-Source Agent. This course isn’t just for developers; we dive into practical use cases for various tasks such as data analysis, content creation, and personalized learning.
How Google SREs use Gemini CLI to solve real-world outages: In this article, we’ll delve into real scenarios that Google SREs are solving today using Gemini 3 (our latest foundation model) and Gemini CLI—the go-to tool for bringing agentic capabilities to the terminal.
Getting started with Gemini 3: Deploy your first Gemini 3 app to Google Cloud Run: In this blog, we will show you how to vibe code your first app—which leverages the Gemini 3 Flash Preview model and deploy it as a publicly accessible URL on Google Cloud Run. Google AI Studio lets you go from idea to app quickly by using natural language to generate fully functional apps using the power of Gemini 3.
Practical guidance: Building with the Secure AI Framework (SAIF) on Google Cloud: We know that security and data privacy are the top concern for executives when evaluating AI providers, and security is the top use case for AI agents in a majority of industries. To help you build AI boldly and responsibly, here’s our guide to developing AI with the Secure AI Framework (SAIF) on Google Cloud.
The truths about AI hacking that every CISO needs to know (Q&A): How will AI boost threat actors? And what can chief information security officers do about it? Google’s Heather Adkins, vice-president, Security Engineering, explores how securing the enterprise is about to change.

How Schrödinger sped up molecular discovery by 4x with Alphaevolve

Tue, 30 Jun 2026 16:00:00 +0000

Computational chemistry researchers have traditionally faced a frustrating trade-off when simulating molecular interactions: use fast classical force fields that sacrifice precision or rely on accurate quantum-mechanical methods that run too slowly on large jobs.

Machine-learned force fields (MLFFs) close that gap by training neural networks on high-fidelity quantum data. When it comes to modern drug discovery and materials design, though, there’s demand for even faster processing speeds to handle massive chemical libraries involved. To overcome such performance constraints, Schrödinger partnered with Google Cloud to deploy AlphaEvolve, an evolutionary AI coding agent developed by Google DeepMind that iteratively generates and refines algorithms to find the most efficient code path overcoming the algorithmic bottleneck.

A collaborative duet with AlphaEvolve

Schrödinger — a leader in developing scientific software for over three decades — identified two critical algorithms within their MLFF training pipeline that limited performance: neighbor list computation and Ewald summation. These algorithms aggregate data from atomic neighbors and calculate long-range potentials, but both became limiting factors in training and inference speed.

Schrödinger's primary technical goal was speeding up AI model training for energy and force calculations. Specifically, they targeted the Ewald summation, a critical but computationally demanding function used in molecular mechanics. The Ewald sum was the main performance constraint in Schrödinger's PyTorch code. It had no established vectorized algorithm and often relied on simple for-loops that ran slowly on large simulations.

By incorporating AlphaEvolve into their models, the system could generate a batched implementation of the Ewald summation using parallel batch matrix multiplication. This would evolve the PyTorch code to outperform existing custom kernels.

Evaluation metrics

Schrödinger used a rigorous multi-layered evaluation framework to confirm the evolved code was both performant and scientifically accurate:

Inverse time (primary metric): The core objective was to maximize throughput by reducing calculation time, from a baseline score of 7.9.
Functional correctness: All evolved programs had to pass a full test suite, including regression tests on complex systems such as disordered water models.
Success rate: This was measured by the share of programs that were both functionally correct and faster than the baseline.

“AlphaEvolve allows us to explore larger chemical spaces faster and more efficiently than ever before. Faster MLFF inference carries real business impact, shortening R&D cycles in drug discovery, catalyst design, and materials development, and enabling companies to screen molecular candidates in days rather than months.” — Gabriel Marques, technical lead of machine learning, Schrödinger

Results: a 4x speedup and breaking bottlenecks

By applying AlphaEvolve, Schrödinger replaced simple for-loops in the Ewald summation code with parallel batch matrix multiplication. This optimization raised the program success rate from less than 1% (40 out of 5,000 evaluations) to more than 60%, while improving the performance metric from the baseline of 7.9 to nearly 30.

Optimizing these foundational algorithms delivered a 4x speedup in both MLFF training and inference. This acceleration lets researchers compress molecular screening timelines and directly benefits several key research areas:

Drug discovery: Identifying viable therapeutic candidates quickly to address urgent medical needs.
Catalyst design: Developing efficient chemical processes for industrial applications.
Materials development: Designing next-generation materials with custom properties for electronics and energy storage.

The next evolution

Schrödinger plans to apply this evolutionary approach to custom GPU kernels to test whether AI-generated code can outperform human-engineered implementations.

Read the full technical paper on AlphaEvolve to learn how evolutionary AI agents optimize scientific codebases, or contact the Google Cloud AI team to discuss accelerating your research workflows.

Build agents even faster with Gemini Enterprise Agent Platform’s fully-managed, remote MCP server

Tue, 30 Jun 2026 16:00:00 +0000

A couple of months ago, we announced that over 50 Google-managed MCP servers are available.

Today, we’ll dive into how to use the Gemini Enterprise Agent Platform remote MCP server to securely connect your external AI agents to the resources inside your Google Cloud environment.

Connect your IDE to Google Cloud

Think of the Agent Platform MCP server as a bridge between your favorite external development tools and your Google Cloud architecture.

If you are building an agent in Antigravity CLI or Claude Code, for example, the Agent Platform MCP server allows that agent to securely interact with your Agent Platform resources. That way, your agent can now easily call models from Model Garden, pull down shared prompt templates, or even manage Notebooks directly within your project – all without ever leaving the IDE.

Quicker time-to-value

The speed at which you deliver value is one of your greatest advantages. But sometimes, connecting external development environments to cloud infrastructure forces a trade-off. Developers want to move fast with minimal setup, while IT teams need strict governance over data access.

The Agent Platform MCP server provides a single, standardized interface for your external agents so you can spend less time writing integration code and more time building useful features. And by running entirely within Google Cloud’s secure infrastructure, it gives you ready-to-use endpoints that protect your data while accelerating your development.

Get the best of both worlds:

Build with open standards: Agents you build outside of Google Cloud stay fully compliant with the open MCP specification. Your external IDEs and frameworks can seamlessly interact with your cloud environment without locking you into a proprietary ecosystem.
Centralized discovery: Catalog your assets with Agent Registry in Agent Platform. It acts as your organization's centralized library, so your teams can securely store, search for, and govern their entire inventory of skills, tools, and other AI capabilities.
Easy access with security and governance: Your connections are protected by default. IT teams can leverage native Cloud IAM Deny policies to ensure external developer frameworks only interact with authorized Google Cloud resources.

How it works: Three simple steps to connectivity

Enable the API: The Gemini Enterprise Agent Platform remote MCP server is automatically enabled when you enable the Gemini Enterprise Agent Platform API within your Google Cloud project.

2. Configure your client: Connect your AI application by following our configuration instructions to point to the remote server.

3. Use toolsets: Access a robust, copyable list of Toolset Endpoints to begin interacting with your Agent Platform resources immediately.

Available toolsets:

MCP Toolsets
Endpoint	Description	Tools
/mcp/generate	Generative AI tools	Core generation features
/mcp/predict	Prediction tools	Inference and raw prediction
/mcp/notebook	Colab enterprise notebook tools	Notebook runtime and execution management
/mcp/endpoints	Endpoint management tools	Lifecycle management for model endpoints
/mcp/models	Model registry tools	Model upload, registry, and deployment
/mcp/tuning	Model fine-tuning tools	Finetuning job management and tracking
/mcp/evaluation	Quality evaluation tools	Automated model quality and instance evaluation
/mcp/prompts	Prompt management tools	Prompt engineering and versioning workflows

Get started today

Visit the Agent Platform page to connect your favorite agent frameworks to the Agent Platform MCP server and start building today.

Bringing speed and strong cost performance to the market with Gemini Omni Flash and Nano Banana 2 Lite

Tue, 30 Jun 2026 16:00:00 +0000

Great creative happens when your tools move at the speed of your ideas. To help you create rich, reliable experiences while reducing regeneration time and costs, we’re adding two new models to Gemini Enterprise Agent Platform.

First, we’re announcing the general availability of Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image). This model is the fastest and most cost-efficient image generation and editing model within the Nano Banana model family. Whether you're rapid-firing ideas, A/B testing ad variations, or powering social apps for millions of users, this model gives you the power to explore, iterate, and scale with speed.

We’re also releasing Gemini Omni Flash in public preview. Grounded in Gemini's real-world knowledge, it powers high-quality video generation and conversational editing. Whether you're executing character or product swaps, performing dynamic style transfers, or adding objects and relighting scenes, this model gives you precise control to edit and refine video assets.

Bring your boldest vision to life with Gemini Omni Flash and Nano Banana.

Both models provide some of the best price-performance among market-leading frontier models for image and video generation and editing.

Gemini Omni Flash: High-quality video generation and editing

Gemini Omni Flash brings conversational video generation and editing directly into your applications. Users can easily embed powerful media models into their agentic workflows to create, remix, and refine video without ever switching platforms.

For comprehensive benchmarking information from Google DeepMind, please visit Gemini Omni.

We built Gemini Omni Flash with a focus across these four key areas:

Conversational editing: Swap characters, relight scenes, or alter angles using natural language while natively maintaining original audio and video tracks.
Multimodal input: Combine text, images, and video inputs to guide video generation. Gemini Omni Flash natively generates audio with every video output, while maintaining character, object, and style consistency.
World knowledge and simulation: It combines an intuitive understanding of physics with Gemini's knowledge of history, science and cultural context, bridging the gap from photorealism to meaningful storytelling.
Text and action synchronization: Render legible text and graphics directly into video, syncing kinetic typography and explainer text with on-screen movements.

Note: Support for audio references, video references, last frame, scene extension and higher resolutions for the Gemini Omni Flash via Gemini Enterprise Agent Platform API will be available soon.

To see the full list of model capabilities and how to integrate it check out the documentation and pricing.

Priced at $0.10 per second of video output, Gemini Omni Flash delivers some of the best price-performance for video generation and editing capabilities on the market.

Businesses using Gemini Omni Flash to build next-gen applications and creative agentic workflows:

“We’re excited to bring Google’s newest models, including Gemini Omni Flash and Nano Banana 2 Lite, to Adobe Firefly, our all-in-one creative AI studio – to help creators move faster from idea to finished content. These new models build on Adobe’s strategy to deliver our pro-grade tools and the industry’s top creative AI models in a connected workflow, giving creators flexibility and control over how they bring their creative ideas to life." – Matt Chotin, Senior Director of Product, Adobe

“The thing that immediately caught my attention was the sheer range of what the Gemini Omni Flash model does. The VFX capabilities surprised me, and looking at it as a producer, that brings in some very interesting possibilities. But the hybrid possibilities are what excite me most. You take the crews you have always worked with in the live-action world, and you bring the breadth of what AI can do now onto the same set.” - Nishant Tahilramani, Creative Director, Invideo

“Through our continued partnership with Google, WPP received early access to the new Gemini Omni Flash’s model and integrated it into WPP Open, our agentic marketing platform. Gemini Omni Flash’s multi-modal capabilities—allowing for seamless image, audio, and video input references—combined with intuitive conversational editing, represent a leap forward for controlled AI production. Teams have tested asset localization, precise product swaps, and dynamic style transfers for clients. We are thrilled to partner with Google Cloud to continually push the boundaries of AI-driven creativity and deliver highly adaptable, intelligent work for our clients.” – Elav Horwitz, Chief Innovation Officer, WPP

Nano Banana 2 Lite: Built for cost and speed

Nano Banana 2 Lite can generate an image in as little as four seconds. You can generate and iterate on design concepts in seconds, taking you from a blank page to the perfect layout instantly.

Nano Banana 2 Lite generates images in as little as four seconds.

Significant improvements over Nano Banana (Gemini 2.5 Flash Image)

Nano Banana 2 Lite blends fast image generation with a significant leap in visual quality and capability compared to our legacy model, Nano Banana. We enhanced core capabilities so you can execute complex tasks at high speeds:

World knowledge: Quickly draft accurate contextual scenes, rough data visualizations, and location-specific mockups.
Character consistency: Maintain character identities and object fidelity across multiple swift generations to easily build out storyboarding tools or embed virtual try-ons for ecommerce.
Quick text and localization: Draft copy on the fly by rendering legible text directly into rapid generations to see how typography works across localized ad variations.

To see the full list of model capabilities and how to integrate it check out the documentation and pricing.

Note: Image generation offers the fastest latency. Image editing may experience slightly higher response time.

Fast, cost-efficient image generation with a significant leap in visual quality and capability compared to our legacy model, Nano Banana

Industry leaders building faster visual experiences with Nano Banana 2 Lite:

“Speed is no longer a limitation. When generation is faster than imagination, creators can stay inside the idea instead of waiting on the tool. Nano Banana 2 Lite brings that feeling into the creative process, letting thoughts move into visuals almost instantly. For Artlist’s users, it means less time staring at a progress bar and more time creating, iterating, personalizing, and moving at the speed of culture.” - Idan Yonas, Director of AI Content & Innovation, Artlist

"Nano Banana 2 Lite is fast and reliable, helping designers explore more ideas to craft unique images on Figma Weave's node-based canvas. It's ideal for rapid iteration while staying in the creative flow." - Itay Schiff, Co-founder and Creative Director, Figma

"We have been testing Nano Banana 2 Lite to power real-time image generation within Manus’s autonomous workflows—from slide decks to web pages. Its speed suits these scenarios well, allowing our AI Agent to iterate on visuals quickly and deliver results in seconds. The image quality is also impressive, coming close to the full Nano Banana 2. We look forward to continuing our partnership and building better experiences together." - Tao Zhang, Co-founder and Chief Product Officer, Manus AI.

Safety and enterprise governance

C2PA content credentials and imperceptible SynthID watermarks are enabled by default to help verify content authenticity for both models.

To handle high-concurrency API requests reliably at scale, the Gemini Enterprise Agent Platform offers provisioned throughput (PT) for Nano Banana 2 Lite starting today. Provisioned throughput for Gemini Omni Flash will be rolling out soon.

Start building today

Embed these image and video generation and editing capabilities into your applications and creative workflows today. Explore these resources to start building:

Try the models: Agent Studio within Gemini Enterprise Agent Platform
API documentation: Nano Banana 2 Lite, Gemini Omni Flash
Access Colab notebooks: Nano Banana 2 Lite, Gemini Omni Flash
Pricing: Agent Platform Pricing for both models
Prompting guides: Nano Banana, Gemini Omni Flash
Gemini Omni Flash Prompting Agent Skills

Cloud CISO Perspectives: How Google Cloud Security uses AI internally

Mon, 29 Jun 2026 16:00:00 +0000

Welcome to the second Cloud CISO Perspectives for June 2026. Today, we’re discussing how we use AI to chart a path to autonomous software development lifecycle security.

As with all Cloud CISO Perspectives, the contents of this newsletter are posted to the Google Cloud blog. If you’re reading this on the website and you’d like to receive the email version, you can subscribe here.

aside_block: <ListValue: [StructValue([('title', 'Get vital board insights with Google Cloud'), ('body', <wagtail.rich_text.RichText object at 0x7fa654f38820>), ('btn_text', 'Visit the hub'), ('href', 'https://cloud.google.com/solutions/security/board-of-directors?utm_source=cgc-site&utm_medium=et&utm_campaign=FY26-Q2-GLOBAL-GCP39634-email-dl-dgcsm-CISOP-NL-177159&utm_content=-&utm_term=-'), ('image', <GAEImage: GCAT-replacement-logo-A>)])]>

Cloud CISO Perspectives: Our path to autonomous SDLC security

By Chris Betz, CISO, and Ruchi Shah, senior director, Security Engineering, Google Cloud

Chris Betz, CISO, Google Cloud

AI has upended the economics of exploiting vulnerabilities, effectively erasing the traditional patching window. To survive this new reality, security requires an autonomous defense.

To counter machine-speed, AI-driven threats, we’ve worked hard to transition Google Cloud’s security posture to an autonomous, proactive model. By embedding specialized AI agents directly into our software development lifecycle (SDLC), we’ve created automated guardrails that protect code at a scale and speed unreachable by human teams — and we’re taking steps to make those same guardrails widely available.

Ruchi Shah, senior director, Security Engineering, Google Cloud

How we designed agentic, secure SDLC architecture

Google Cloud deploys modular, interconnected AI agents across every stage of the software lifecycle to continuously harden products from code ingestion to production.

1. Design, review, and gate

Historically, launch intakes and threat modeling were manual bottlenecks. Today, Google Cloud engineering teams route product launches through an agent-based security review pipeline.

Agents cross-reference designs against a continuous control catalog of more than 200 rigorous security requirements. High-risk indicators are automatically triaged and flagged for human engineering intervention, while a dynamic product dossier updates in real-time to replace static threat models.

Google Cloud has embedded agentic capabilities across the entire SDLC flow to continuously harden products end-to-end.

2. Centralized AI code scanning and the Mantis framework

Naive, decentralized AI code scanning suffers from sloppiness, frequently hallucinating bugs and yielding true-positive rates under 7%. To solve this, we built Mantis, our core multi-agent orchestration framework designed specifically for scalable, context-aware repository analysis.

The core skills at the heart of Mantis are now open source to demonstrate the fundamental concept. We have a more full-fledged version running internally and securing our customers.

Mantis eliminates brute-force code ingestion by constructing a hierarchical security summary tree. By condensing individual files into directory and root-level summaries, Mantis reduces token overhead by over 85% while preserving critical structural context across massive repositories.

The architecture relies on a highly-coordinated workflow across new agents and existing technologies:

Strategist agent: Evaluates the high-level code structure, threat models, and dependency graphs to isolate risky architectural patterns, establishing a prioritized global plan of targeted investigation tasks.
Research agents: Acting as specialized domain investigators, these agents use internal code searches to drill into raw source files, examining data tracking, control flows, and sanitization logic.
Deduplicator, reviewer, and critic agents: Sanitize findings to filter out noise and eliminate false positives.
Reproduction sandbox: Automatically runs AI-generated proof-of-concept exploits in an isolated, emulated environment to verify real-world exploitability before alerting developers.

3. Self-healing fuzz testing

While code scanning provides breadth, dynamic fuzz testing uncovers deep runtime vulnerabilities. However, writing and maintaining fuzz harnesses are often a significant engineering bottleneck.

Our autonomous, multi-agent engine eliminates manual intervention:

Context and Drafting agents synthesize product logic and existing unit tests to author initial fuzzing harnesses.
Building and Testing agents execute the code and feed real-time compiler and linker errors into a Hallucination Cleaner agent, which acts as an automated mechanic to repair broken dependencies and build configurations.
Quality Analyzer agents monitor runtime execution, actively adjusting inputs to bypass code blockers and penetrate deeper into complex, stateful APIs.

4. The unified AI patching pipeline

Finding thousands of vulnerabilities at scale can create a dangerous remediation backlog without proper planning. To close the exposure window, our discovery tools route findings directly into an autonomous remediation pipeline:

The Reproduce agent replicates the crash in the sandbox.
The Bug Context agent maps the failure execution path.
The Patch agent generates a targeted code fix.
The Evaluation agent runs a rigorous regression loop (that re-compiles code and executes tests) to ensure the patch is safe. Only fully-validated fixes are submitted to a human reviewer.

5. Autonomous and secure posture management

Post-launch, we maintain security integrity with an autonomous security posture management (ASPM) system. By converting our security standard catalog into programmable skills files, the ASPM system continuously checks production systems for configuration drift, automatically triggering agentic remediation when a violation occurs.

Continuous augmentation via self-reflection

Stateless AI systems repeatedly fall into the same logical traps, such as attempting to fix bugs inefficiently and hallucinating about non-existent code. Our framework solves this by introducing a post-hoc self-reflection loop. After a workflow concludes, a dedicated reflection agent analyzes execution logs, tool histories, and human feedback.

Successful trajectories and design patterns are permanented into a global knowledge store. When future agents spin up, this intelligence is injected directly into their context window, creating a compounding-interest effect on our security engineering. This approach has helped us to improve both the vulnerability fix success rate and efficiency.

Moving toward immune software

Google Cloud's internal journey demonstrates that protecting software at AI-scale requires a fundamental paradigm shift from human-dependent checklists to proactive multi-agent orchestration. By pairing open-source tooling like Mantis with autonomous, self-healing execution loops, we are pioneering a future of "immune" software development — where applications continuously discover, validate, and patch their own weaknesses in real-time.

You can learn more about how we use Mantis and other tools to find and fix vulnerabilities at machine-speed here.

aside_block: <ListValue: [StructValue([('title', 'Learn something new'), ('body', <wagtail.rich_text.RichText object at 0x7fa654f38610>), ('btn_text', 'Watch now'), ('href', 'https://www.youtube.com/watch?v=C1wEjzOHh7Y'), ('image', <GAEImage: Cloud-CISO-Perspectives-logo-A>)])]>

In case you missed it

Here are the latest updates, products, services, and resources from our security teams so far this month:

Verifiable trust in the AI era: What’s new in Confidential Computing: To help further strengthen verifiable privacy in cloud AI deployments, here’s our latest Confidential Computing innovations. Read more.
Choice, compliance, and collaboration: Europe’s path to open digital sovereignty: Our Sovereign Cloud solutions are designed to meet Europe's tiered compliance requirements at every level. Read more.
How AI Is rewriting the SecOps playbook: With adversaries operating at machine speed, defenders must prioritize speed, automation, and continuous decision-making. Read more.
Google named a Leader in IDC MarketScape SIEM 2026 Vendor Assessment: We are proud to announce that Google has been named a Leader in the 2026 IDC MarketScape for worldwide SIEM platforms. Read more.
Announcing the Wiz Runtime Sensor for Windows: Wiz pairs real-time threat detection with a memory-safe architecture that scales efficiently to protect your essential cloud infrastructure. Read more.
New VPC Service Controls updates can help secure agents: Designed for agentic workloads, new capabilities in VPC Service Controls can help establish a network-level, destination-based perimeter. Read more.
Bug hunting on Gemini Spark: Gemini Spark brings a persistent agent to the Gemini App. Learn how to approach security testing for this new paradigm and focus on high-impact bugs. Read more.

Please visit the Google Cloud blog for more security stories published this month.

aside_block: <ListValue: [StructValue([('title', 'Join the Google Cloud CISO Community'), ('body', <wagtail.rich_text.RichText object at 0x7fa654f387f0>), ('btn_text', 'Learn more'), ('href', 'https://rsvp.withgoogle.com/events/google-cloud-ciso-community-interest-form-2026?utm_source=cgc-blog&utm_medium=blog&utm_campaign=FY25-Q1-global-GCP30328-physicalevent-er-dgcsm-parent-CISO-community-2025&utm_content=cisop_&utm_term=-'), ('image', <GAEImage: GCAT-replacement-logo-A>)])]>

Threat Intelligence news

China-nexus threat actor targets medical community for cross-sector research: Google Threat Intelligence Group (GTIG) has identified a sophisticated campaign attributed to UNC6508, a People's Republic of China (PRC)-nexus threat actor, targeting the North American academic, medical, and military research community, that went undetected for more than a year. Read more.
ShinyHunters targets education sector with Oracle PeopleSoft exploit: Mandiant and GTIG have identified an active compromise and extortion campaign attributed to UNC6240 (ShinyHunters) targeting Oracle PeopleSoft application infrastructure. Read more.
Zero-day exploitation in Cisco Catalyst SD-WAN Manager: Mandiant has identified a threat actor targeting a vulnerability in Cisco Catalyst SD-WAN to escalate privileges from a compromised administrative account to root-level access. Read more.

Please visit the Google Cloud blog for more threat intelligence stories published this month.

Now hear this: Podcasts from Google Cloud

Cloud Security Podcast: How Google Cloud uses LLMs to defend billions of users: Google Cloud CISO Chris Betz discusses AI Threat Defense, and emphasizes shifting security practices earlier in the development lifecycle through human-AI collaboration. Listen here.
Cloud Security Podcast: To couple or decouple SIEM: Alex Hurtado, director, Detection Engineering, Scanner, and Christopher Witter, DNR lead, Dropbox, debate the merits of centralized versus decentralized SIEM architectures. Listen here.

To have our Cloud CISO Perspectives post delivered twice a month to your inbox, sign up for our newsletter. We’ll be back in a few weeks with more security-related updates from Google Cloud.

Synthesize the big picture and analyze trends with BigQuery's AI.AGG function

Mon, 29 Jun 2026 16:00:00 +0000

We recently announced the preview of the BigQuery AI.AGG() function. With AI.AGG(), you can use natural-language instructions within a single line of SQL to summarize or synthesize information over millions of rows of unstructured or even multimodal data.

While BigQuery already offers powerful AI functions that help you analyze individual rows of data, analyzing unstructured data at scale requires a different approach. AI.AGG() lets you ask questions from unstructured data such as logs and documents, for example:

What are the top three feature requests among the negative product reviews?
What kind of errors are users seeing most frequently, and how should I start investigating them?
In which specific scenarios is our automated agent consistently failing to resolve customer issues?

In this post, we'll dive deeper into the AI.AGG() function and look at a few of the use cases that it unlocks, including how it can be used in combination with BigQuery’s other managed AI functions for complex, intelligent data analysis.

Analyzing system logs with `AI.AGG()`

A great example of the power of AI.AGG() is analyzing system logging. Log messages, warnings, errors, and stack traces can contain extremely useful information for improving your service, but it can be time- and labor-intensive to investigate them manually — especially if you operate at scale and have thousands of them to review.

With AI.AGG(), you can easily analyze many logs at once, grouping and prioritizing them to decide which ones to dig deeper into first. In fact, our BigQuery engineering team used this exact approach while developing AI.AGG() — using the function to help identify edge cases related to input handling for the feature itself!

To demonstrate this, let’s analyze a public dataset of Apache Spark standard INFO logs available from Loghub. Often, clusters can run into issues like memory thrashing, clock drift, or broadcast bottlenecks without ever throwing a FATAL error. You can use AI.AGG() to analyze these seemingly normal logs for hidden inefficiencies. You can load the sample data file into BigQuery using any of the supported methods, such as the UI, CLI, or client libraries. The following example assumes you’ve loaded the log file into a dataset called bq_logs_demo and table named spark_logs_unstructured.

Notice how we construct the prompt here. We explicitly give the model permission to say "everything is fine," which prevents it from hallucinating errors, while instructing it to hunt for specific anomalies:

code_block: <ListValue: [StructValue([('code', "SELECT\r\n Component AS spark_component,\r\n COUNT(*) AS log_count,\r\n AI.AGG(\r\n Content,\r\n 'Analyze these Spark system INFO logs. Provide a 2-sentence summary: First, describe the normal operation of this component. Second, explicitly identify any hidden inefficiencies, latency spikes, repeated retries, or unusual patterns.'\r\n ) AS performance_analysis\r\nFROM\r\n `bq_logs_demo.spark_logs_structured`\r\nGROUP BY\r\n Component\r\nORDER BY\r\n log_count DESC;"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654ec1670>)])]>

You can see in these results that AI.AGG() successfully acknowledges the "operating normally" messages while surfacing the critical diagnostic insights:

The query results pane showing the insights generated by AI.AGG() over the logs dataset.

Extracting categories from unstructured text and image data

Now, let’s look at some more use cases that demonstrate the flexibility of AI.AGG(), using one of BigQuery’s public datasets, cymbal_pets, a fictional pet supply shop. It includes a catalog of products carried by the store, with unstructured data like product names, descriptions, and images, making it a great example of the power of AI functions for handling unstructured data.

For example, let’s say you want to categorize the products in the dataset. The first hurdle in this case isn't applying labels to your products, but discovering what categories exist across the product catalog. With AI.AGG(), you can ask the model to analyze the raw product names and descriptions to identify the overarching categories for you.

code_block: <ListValue: [StructValue([('code', "-- Identify categories of products from product name and description\r\nSELECT\r\n AI.AGG(\r\n ('Product: ', product_name, ' - Description: ', description),\r\n 'What are the major categories of these products?' \r\n ) AS category_description\r\nFROM\r\n `bigquery-public-data.cymbal_pets.products`;"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654ec1880>)])]>

This query returns a simple plaintext list of categories:

The plaintext result of categories determined by AI.AGG() over our products dataset.

This initial query is great for discovery, but a simple plaintext string isn't enough to build a reliable, automated data pipeline. To actually tag your data, you need to instruct AI.AGG() to return a structured format, like a JSON array. Then, you can use the structured categories as a parameter within another AI function, AI.CLASSIFY(), to actually label each product with its category.

The following SQL statement completes each of these steps in one script:

code_block: <ListValue: [StructValue([('code', "-- 1. Declare a variable to hold the array of categories\r\nDECLARE generated_labels ARRAY<STRING>;\r\n\r\n-- 2. Create a dataset to store the results\r\nCREATE SCHEMA IF NOT EXISTS categorized_cymbal_pets;\r\n\r\n-- 3. Generate the JSON string with AI.AGG and extract it into the variable\r\nSET generated_labels = (\r\n SELECT \r\n JSON_VALUE_ARRAY(\r\n AI.AGG(\r\n ('Product: ', product_name, ' - Description: ', description), \r\n 'Identify the major product categories. Return exactly one valid JSON array of strings. Do not include markdown code blocks, backticks, or conversational text.'\r\n )\r\n )\r\n FROM `bigquery-public-data.cymbal_pets.products`\r\n);\r\n\r\n-- 4. Feed the variable directly into AI.CLASSIFY\r\nCREATE OR REPLACE TABLE `categorized_cymbal_pets.categorized_products` AS (\r\nSELECT \r\n product_name,\r\n description,\r\n AI.CLASSIFY(\r\n ('Product: ', product_name, ' - Description: ', description),\r\n generated_labels\r\n ) AS assigned_category\r\nFROM \r\n `bigquery-public-data.cymbal_pets.products`\r\n);"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654ec1850>)])]>

You can now view the resulting table, which includes an assigned_category:

A preview of the categorized_products table which includes the new assigned_category column created by AI.AGG() and AI.CLASSIFY().

If you look closely at the intermediate table, you'll notice the structured categories changed slightly from the initial plaintext results. This happens for two reasons: First, LLMs are nondeterministic, meaning that they don't always give the exact same response to the same prompt. Second, the prompt was adjusted to accommodate the new output structure.

The returned product categories are structured as JSON by AI.AGG() as requested as part of the prompt.

With the table now labeled by category, you can group by the categories to do traditional SQL aggregation, or use AI.AGG() to consider each category separately.

For example, the following query fetches traditional metrics (like row counts) right alongside a synthesized AI summary of what those specific grouped products have in common:

code_block: <ListValue: [StructValue([('code', "-- Synthesize insights grouped by our newly assigned categories\r\nSELECT \r\n assigned_category,\r\n COUNT(*) AS item_count,\r\n AI.AGG(\r\n ('Product: ', product_name, ' - Description: ', description),\r\n 'Write a concise, one-sentence summary describing the common characteristics or purpose of the products in this category.'\r\n ) AS category_summary\r\nFROM \r\n `categorized_cymbal_pets.categorized_products`\r\nGROUP BY \r\n assigned_category\r\nORDER BY \r\n item_count DESC;"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654ec1820>)])]>

Query results showing analyzing with AI.AGG() alongside more traditional SQL methods.

Unstructured data isn't limited to text. Because AI.AGG() natively supports multimodal inputs, you can return aggregated insights directly from image files.

The cymbal_pets Google Cloud project also contains a Cloud Storage bucket full of product photos. By creating an external object table, you can securely pass the image URIs directly into AI.AGG() and ask the model to summarize the visual content of the entire collection.

code_block: <ListValue: [StructValue([('code', "-- Summarize content of images in the object table\r\nSELECT\r\n AI.AGG(\r\n STRUCT(OBJ.GET_ACCESS_URL(ref, 'r')),\r\n 'What are the major categories of these images?'\r\n ) AS category_description\r\nFROM\r\n `bigquery-public-data.cymbal_pets.product_images`;"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654ec1c40>)])]>

Query results showing AI.AGG() surface product categories by analyzing the product images located in Google Cloud Storage.

How AI.AGG() works and best practices

To use AI.AGG() effectively in your own environment, it helps to understand how it processes data behind the scenes. Here’s what you need to know about context windows, error handling, and optimizing your pipelines.

1. Context windows and multi-level aggregation
LLMs have a specific context window and can have a hard time handling massive amounts of input. AI.AGG() solves this problem by automatically dividing your input rows into batches, aggregating those batches, and then aggregating the results of those batches into a final answer. This means you don’t have to worry about manually managing the context window when passing in large numbers of rows. Note that AI.AGG() won’t split up a row of data across batches, so make sure that each individual row is smaller than the context window, to avoid the row being skipped. Many smaller rows will give AI.AGG() more flexibility with how to batch each row.

2. Token usage with multi-level aggregation
Because AI.AGG() uses a multi-level aggregation structure, the total input tokens sent to the model may be higher than the raw tokens in your starting table (depending on how many rounds of aggregation are required). As a best practice, always reduce the number of input tokens by using LIMIT or pre-filtering your data upstream before passing it to AI.AGG().

3. Specifying your model endpoint
If you don’t specify a model endpoint, AI.AGG() will default to a recent model. However, for production pipelines, you often want explicit control:

Short-form names: You can use a short-form endpoint (e.g., gemini-2.5-flash), in which case AI.AGG() will use that model in the query execution region:

code_block: <ListValue: [StructValue([('code', "AI.AGG(\r\n input_data,\r\n instructions => 'Your instructions here.',\r\n endpoint => 'gemini-2.5-flash' \r\n)"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654ec1970>)])]>

Fully-qualified names: If the query execution region doesn’t support your desired model, or you prefer to use a global or multiregional endpoint, provide the fully qualified model name:

code_block: <ListValue: [StructValue([('code', "AI.AGG(\r\n input_data,\r\n instructions => 'Your instructions here.',\r\n endpoint => 'https://aiplatform.googleapis.com/v1/projects/[YOUR_PROJECT]/locations/global/publishers/google/models/gemini-3.5-flash'\r\n)"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654ec18e0>)])]>

4. Input and output modalities

Inputs: AI.AGG() supports text (via strings or references to text files) and image data. It also supports arrays of these types, though you should refer to the known issues documentation for edge cases regarding arrays of images.
Outputs: The function will always return a string. While you can prompt the model in your instructions to format the output as JSON or Markdown, keep in mind that the database engine does not strictly enforce this. Multimodal output (e.g., generating an image) is not currently supported.

5. Treatment of NULLs
AI.AGG() automatically skips NULL input rows without processing them. However, you must be careful when passing structured data. Like other BigQuery AI functions, AI.AGG() concatenates STRUCT fields similarly to the standard CONCAT() function. This means if even one field within your STRUCT is NULL, the entire row is treated as NULL and will be skipped.

Let's revisit our first categorization query. What if several rows of our products table are missing their description? Because of the NULL concatenation rule, those rows would be silently dropped from the analysis entirely. Here is how we can use IFNULL() to provide a fallback string, guaranteeing that every product is taken into account even if its description is blank:

code_block: <ListValue: [StructValue([('code', "-- Identify categories of products from product name and (optional) description\r\nSELECT\r\n AI.AGG(\r\n ('Product: ', product_name, ' - Description: ', IFNULL(description, 'No description provided')),\r\n 'What are the major categories of these products?' \r\n ) AS category_description\r\nFROM\r\n `bigquery-public-data.cymbal_pets.products`;"), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654ec1f40>)])]>

6. Error handling
If AI.AGG() receives invalid input, or encounters an error during LLM processing, it will attempt to provide partial results. Rows containing invalid input or which were rejected by the LLM model will not be considered in the final results.

You can review exactly how many rows failed to process by checking your BigQuery job statistics, exactly as you would for scalar managed AI functions like AI.IF().

information showing an example of Gen AI function error details.

Give it a try!

These are just a few examples of the ways AI.AGG() can help analyze unstructured data. The AI.AGG() function is in preview in BigQuery now, so it’s available to all BigQuery users. Try it out on your own use cases!

You may also be interested in checking out BigQuery's other managed AI functions, AI.CLASSIFY(), AI.IF(), and AI.SCORE(), as well as general-purpose functions like AI.GENERATE(). We look forward to seeing what you build with them.

Securing agentic AI with perimeter guardrails: What's new in VPC Service Controls

Fri, 26 Jun 2026 18:00:00 +0000

As enterprises scale autonomous AI agents into production, enabling safe innovation requires robust architectural guardrails. AI agents connect across tools and datasets, so it’s essential to establish clear network-level boundaries for comprehensive data protection.

To help organizations confidently deploy these workflows, we recommend VPC Service Controls (VPC-SC) to establish an essential network-level, destination-based perimeter. Today we’re announcing several new capabilities specifically designed for agentic workloads.

What's new in VPC Service Controls

Designed to enhance AI security, the new capabilities we’re announcing today strengthen boundaries enforced by VPC-SC.

The capability updates include:

Agent identity in directional rules: Enforcing least-privilege access requires treating agents as first-class identities. You can now add agentic identities directly to service perimeter ingress and egress rules using standard Identity and Access Management (IAM) principals.

A single principal maps to an individual agent, while a principalSet maps to a broader collection of agents. PrincipalSets lets administrators apply consistent, auditable access policies across agent fleets. If an agent is compromised, you can immediately revoke its access at the network perimeter.

Granular control with model context protocol (MCP) attributes: As MCP becomes the standard integration layer for agentic systems, the ability to enforce policy at the tool level is critical. VPC Service Controls now support conditional access rules based on specific MCP attributes, including mcp.toolName, mcp.method, and mcp.tool.isReadOnly.

For example, you can grant an agent read access to a Workspace MCP server while explicitly denying its ability to send emails.

Securing the Gemini Enterprise Agent Platform: The Gemini Enterprise Agent Platform provides a comprehensive foundation for production-grade agent deployments. VPC Service Controls is now natively integrated with Agent Platform. When you include Agent Platform as a protected service within a VPC-SC perimeter, the system automatically blocks all public internet access to the Agent Platform instance — enforcing a secure boundary without additional configuration overhead.

"At Mercado Libre, VPC Service Controls serve as an essential, foundational layer of our security architecture. By building a strong perimeter enforcement across hundreds of Google Cloud projects in our organization, we established robust network-level security controls with VPC-SC, ensuring all our data remains protected in our cloud environment," said Juan Pablo Boschi, project lead at Mercado Libre.

Defining a layered approach to enterprise AI security with VPC-SC

Securing an autonomous agent requires a layered approach. Identity, network, and resource controls each target a distinct threat vector.

Identity controls: IAM and Principal Access Boundaries (PAB) focus on "who" can access specific resources. By enforcing strict least-privilege principles for agent identities, you help ensure that autonomous workloads only have the permissions necessary for their specific objectives.
Network controls: Next-generation network firewalls and VPC Service Controls define a robust data perimeter on top of your infrastructure, governing the flow of information across boundaries and preventing data exfiltration.
Resource controls: Organization Policy and other resource-level guardrails set broad, immutable constraints on how resources can be configured and used, preventing risky configurations by default.

While identity and network controls effectively secure the front door, VPC Service Controls provide a critical destination-based defense. In the probabilistic world of autonomous agents, VPC-SC is the control that focuses on the "how” and "where" of the agent’s network and operations, in addition to the “who”.

Defending against the unique attack vectors

Unlike traditional applications, an AI agent's input can inadvertently prompt it to execute an unintended command or action. If an agent is successfully compromised — whether driven by malicious prompts, tool manipulation, or malicious insider commands — VPC Service Controls serves as a critical network safety net.

To illustrate how this network boundary defends against industry-standard risks as mapped by the OWASP Top 10 for LLM Applications, here are three real-world threat vectors where VPC Service Controls can help supplement identity-based controls to prevent data exfiltration.

Exfiltration prevention via indirect prompt injection (OWASP ASI01): A malicious actor could attempt to embed a hidden prompt asking an agent to summarize internal data and transmit it to an unauthorized user. If the hijacked agent has IAM permissions, IAM detects no anomaly.

However, when the agent tries to send that data to an external webhook, VPC-SC blocks the API-layer transfer because the destination is outside the defined perimeter.

Guardrail for tool misuse (OWASP ASI02, ASI08): Prompt hijacks can lead agents to chain tools maliciously, such as sending internal directory data to an external service. By enforcing a VPC-SC perimeter around sensitive assets, you prevent misbehaving agents from bridging data across isolated trust zones.
Neutralizing insider threats (OWASP AS103): Attackers can command a data-processing agent to perform a direct cloud-to-cloud copy from a BigQuery dataset to an unauthorized project. While network firewalls see legitimate HTTPS traffic to BigQuery, and IAM sees an authorized service account, VPC-SC evaluates the destination resource. Since the destination project is outside the enterprise perimeter, the system immediately denies the API request.

VPC Service Controls acts as a perimeter to block data exfiltration attempts from a compromised agent, even if the agent has valid IAM credentials.

Data protection for the autonomous agent world

Perimeter security has evolved from a recommended best practice in the deterministic application and workload centric age to an absolute requirement for the era of autonomous AI agents. VPC-SC provides the necessary control over data movement that IAM cannot address alone.

In an era where agents interpret prompts as code, VPC-SC becomes the mandatory safety net for enterprise data. Pairing the mapping capability of IAM with the rigid data perimeters of VPC-SC lets organizations securely build agentic innovation while maintaining an absolute guardrail against exfiltration.

To learn more, you can explore VPC-SC resources here.

Verifiable, private AI: Google Cloud expands Confidential Computing frontiers

Tue, 23 Jun 2026 16:00:00 +0000

Protecting sensitive data used with AI is a critical part of our commitment to providing advanced and secure cloud infrastructure. Confidential Computing cryptographically protects data in use in hardware-based Trusted Execution Environments (TEEs) with verifiable data integrity.

We are thrilled to share our latest Confidential Computing innovations across our hardware ecosystem that help further strengthen verifiable privacy in cloud AI deployments.

Confidential AI at global scale

By scaling our Confidential AI capabilities globally, we help ensure that AI inference and fine-tuning workloads can run with enforceable privacy guarantees.

Democratizing Confidential AI: Confidential G4 VMs with NVIDIA RTX PRO 6000 Blackwell GPUs in preview

We are excited to announce a landmark moment for accessible Confidential AI at global scale: Confidential VMs and Confidential GKE Nodes on the accelerator-optimized G4 machine series, featuring NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs.

What makes this a game-changer is its global scale and flexibility. Confidential G4 is available in every Google Cloud region that the standard G4 is available, across multiple consumption models including On Demand, Reservations, DWS Flex Start, and Spot/Preemptible.

"As organizations scale AI across multiple infrastructure environments, maintaining privacy and control over data and execution becomes increasingly challenging. Google Cloud Confidential G4 VMs powered by NVIDIA RTX PRO 6000 Blackwell GPUs are a meaningful addition to the expanding Confidential AI infrastructure ecosystem. As AI workflows now span agents, data sources, and infrastructure boundaries, Super Protocol provides a consistent Confidential AI operating model across Google Cloud Confidential VMs, other clouds, and on-premises environments — abstracting away confidential computing complexity and allowing teams to focus on AI outcomes," said Yulia Gontar, COO, Super Protocol.

Powered by 5th Generation AMD EPYC Turin CPUs leveraging AMD SEV, the G4 machine series with NVIDIA RTX PRO 6000 Blackwell GPUs activates robust hardware-based security. This architecture helps ensure that sensitive data is protected during processing inside the TEE, while also encrypting data as it travels between the CPU and GPU.

"GCP's Confidential G4 VM was the obvious choice for Vertebrae because privacy and security are non-negotiable for our customers. Our product processes sensitive work discussions, so we need to support hardware-signed attestation that both CPU and GPU are running in a trusted execution environment. Using confidential computing on Google Cloud lets us deliver the frontier of AI privacy in the cloud," said Andy Qin, CEO, Vertebrae.

With Confidential G4, you can unlock AI inference, fine-tuning, HPC, and use cases involving highly restricted data, sensitive models, or private prompts, all with minimal performance impact. Get started with Confidential G4 VMs and Confidential G4 GKE Nodes.

Enabling end-to-end private inference: Open-source Prompt Encryption SDKs

Even as we make Confidential AI accessible, we understand that protecting sensitive data in AI workloads goes beyond securing the model execution environment. The prompts and responses themselves can contain highly-confidential information. To provide cryptographic protection for the entire inference lifecycle, we are happy to announce the open-source launch of our Prompt Encryption SDKs, now available on GitHub.

This toolkit helps you establish an end-to-end secure channel for your AI inference workloads, ensuring that prompts are cryptographically protected from the moment they leave the client until they are processed in the TEE; model responses are similarly protected all the way back to the client.

Prompt and response encryption using Prompt Encryption SDK.

The Client SDK is integrated into the client application and works in tandem with the Server SDK integrated into the inference server running in the TEE. Once the SDKs have been used to establish an attested TLS session, the client can be confident that the server is running an authorized workload within a verified Confidential Computing environment.

The client app can then send encrypted prompts to the inference server, knowing that only this server will be able to decrypt and process it in the TEE. Once the server has a response ready, it sends it back via the same encrypted channel to the client app.

You can get started today with the GitHub repository and the Codelab.

Enabling Apple Private Cloud Compute on Google Cloud

Our commitment to privacy is deeply exemplified by our collaboration with Apple to expand Private Cloud Compute (PCC) on Google Cloud.

We are proud to collaborate with Apple to extend Apple’s privacy and security commitments to PCC on Google Cloud. Our platform supports Apple’s PCC privacy commitments with a layered security approach built upon Google Cloud’s infrastructure. This includes leveraging Google Cloud Confidential Computing with Intel TDX, NVIDIA Confidential Computing with NVIDIA Blackwell GPUs, our Titanium security architecture with the Titan chip, and a co-engineered open-source host stack to ensure verifiable transparency.

Together, these technologies help Apple PCC on Google Cloud meet stringent requirements for data protection and user privacy. To dive deeper into this collaboration, read our blog post: Powering the next era of Confidential AI.

Advancing confidential foundations

Google Cloud is committed to making Confidential Computing capabilities broadly available across our infrastructure. Our goal is to integrate hardware-based security features deeply into our foundational compute offerings, allowing customers to enhance data protection without compromising performance or operational flexibility.

Bringing Intel Trusted Domain Extensions (TDX) to the C4 machine series

Confidential VMs with Intel TDX on the C4 machine series will be available in preview soon.

Powered by the latest 6th Generation Intel Xeon processors, this integration offers a significant leap in compute density and performance for data-intensive workloads. By using Intel TDX, C4 instances create hardware-isolated Trust Domains (TDs) that protect sensitive applications and data from the underlying host and hypervisor.

This architecture provides confidentiality and privacy while enabling remote attestation so you can cryptographically verify the environment before processing sensitive data. Best of all, you can turn Confidential Computing on with a few clicks and no code changes.

Expanding Live Migration capabilities

Running mission-critical production environments requires high availability and continuous uptime, even during scheduled cloud maintenance.

Live Migration on C3D-based Confidential VMs is now generally available. This capability allows Google Cloud to perform planned hardware maintenance without interrupting workloads or exposing encrypted guest memory, ensuring seamless uptime for long-running confidential applications.

Enhancing trust and collaboration: Innovations in Confidential Space

Confidential Space is a Confidential Computing environment designed to enable secure multi-party computation and data sharing. It allows organizations to collaborate on sensitive data, such as for joint machine learning or data analytics, without revealing the data to each other or to Google Cloud.

“Google Cloud Confidential Space allows us to provide financial institutions with security guarantees similar to or better than an on-prem service," said Olivier Richaud, vice-president, Platforms and Site Reliability Engineering, Symphony. "Transitioning such security and privacy-sensitive customers to a cloud-based SaaS service would have been impossible without the power of Confidential Computing.”

A key design principle of Confidential Space is to remove the workload operator from the trust boundary, providing cryptographic assurance that only the authorized, attested workload can access the data.

“As AI systems increasingly act on behalf of consumers in financial services, trust in how data is processed becomes paramount. At Sahamati, we see Google Cloud Confidential Space as a foundational technology for enabling privacy-preserving AI in India’s Open Finance ecosystem, creating the trust needed for innovation while maintaining strong security and accountability guarantees,” said Kiran Gopinath, chief innovation officer, and Head, Sahamati Labs.

Our new advancements for Confidential Space provide greater flexibility and stronger assurances. Key updates include:

Independent Verification: Integration with Intel Trust Authority

We are pleased to announce that Intel Trust Authority (ITA) is now generally available as an independent attestation verifier service for Confidential Space.

This integration enables organizations to independently verify the integrity of the Confidential Space environment using Intel’s hardware-rooted attestation before encryption keys are released to workloads. By decoupling attestation verification from the cloud service provider, customers benefit from enhanced transparency, stronger assurance, and a more robust trust model.

"With Confidential Computing woven into our core infrastructure, Google Cloud and Intel are making hardware‑rooted security and independent attestation part of the default fabric of modern compute. From Intel TDX‑powered C4 Confidential VMs running production workloads, to Confidential Space with Intel Trust Authority — now generally available — enabling verifiable multi‑party collaboration, customers can now encrypt, verify, and scale their most sensitive AI and data workflows without rewriting applications or compromising performance, even in the most demanding regulatory environments,” said Anand Pashupathy, general manager and vice-president, Intel Product Assurance and Security (IPAS), Intel Corporation.

Accelerating secure collaboration: Confidential Space with H100 GPU support

To power secure multi-party AI and machine learning, Confidential Space support for NVIDIA Hopper GPUs is now generally available. This can help multiple parties pool their data for training and inference within a Confidential Space environment, using the power of Hopper GPUs, while ensuring that their individual data remains protected from other participants and from Google Cloud.

Confidential Space unlocks use cases like federated learning on sensitive datasets, and building joint models without centralizing data.

“Confidential GPU support in Google Cloud Confidential Space removes one of the biggest barriers to adopting secure AI: the tradeoff between protecting sensitive workloads and achieving production-grade performance," said Adi Hirschtein, VP Product, Duality. "For Duality customers in healthcare, financial services, and government, this enables federated learning, confidential AI, and encrypted RAG workflows to run on sensitive data at scale while keeping data and models protected throughout processing.”

Next steps

Confidential Computing is becoming an essential layer of cloud computing in the AI era. Explore our expanding portfolio of Confidential VMs, accelerated hardware, and open-source tools to see how you can enable secure collaboration and private AI innovation within your organization.

To learn more, join us at the Confidential Computing Summit on June 23 and 24, 2026.

How growing UK midsize businesses are building in the AI era

Wed, 17 Jun 2026 08:00:00 +0000

The UK’s 5-million-plus small and midsize businesses and enterprises (SMBs) are the backbone of our economy. Today, we’re seeing these critical businesses begin to put AI to work, to operate more efficiently, move faster, and ultimately deliver better outcomes for their customers.

This shift is driven by tangible day-to-day results. According to recent research from Enterprise Nation published in partnership with Google, 71% of AI adopters surveyed in the UK say the technology helps them save time on routine tasks, and 64% report a direct boost in productivity. On top of this, AI-enabled productivity tools (like Google Workspace with Gemini) are delivering a 20% boost in productivity for SMBs, which effectively hands them back one full working day every single week.

At Google Cloud, we have a front row seat to this shift: SMBs have long utilized platforms like Google Workspace, and today they’re transforming with Google’s AI platform and models. In fact, we’ve seen the number of UK-based SMBs using Google Cloud AI nearly double year-over-year. This includes our Gemini models and products like Gemini Enterprise and AI Studio, which are helping SMBs do things like:

Roll out better customer support systems to help escalate and resolve customer support calls more quickly.
Automate repetitive actions in areas like payroll and accounting.
Help more employees understand and leverage data at work — even those not trained as data analysts.
Rapidly create and implement new designs for marketing collateral.
Help more people build their own AI agents to help them in their everyday jobs.
Conduct complex research projects at a speed and price point previously unavailable.

At today’s Google Cloud London Summit, we’re showcasing a number of innovative SMB customers who are actively using our AI tools to transform how they work, including companies who have recently expanded their work with us:

Neural Alpha, a sustainability fintech company, is using Gemini models to read unstructured environmental and corporate sustainability reports to automatically find and organize thousands of key facts, cutting months of slow, manual research down to a fraction of the time.
Sep 2, a digital security provider, uses Gemini Enterprise to deploy autonomous AI agents for 24/7 threat monitoring — accelerating incident detection and quickly neutralizing security threats for its customers.
Sunhouse, a strategic brand design agency, uses Gemini Enterprise to easily find archived design work stored on Google Drive, enabling its teams to spend less time hunting for files and more time growing its business with global brands.
Terrapinn, a global B2B events company, is transforming its operations by leveraging Gemini models, NotebookLM, Looker, and BigQuery to turn manual tasks into automated workflows, accelerating how its teams design, market, and deliver world-class conferences.
VoCoVo, a telecommunications provider, is integrating Google Cloud AI across its systems to turn isolated data into actionable intelligence and build autonomous workflows, streamlining routine operations so their team can focus on high-impact innovation.

Empowering Your Team: AI Upskilling Resources for Growing British Businesses

To help midsize teams maximize their impact and confidently navigate the modern AI landscape, we’ve developed a suite of dedicated, no-cost upskilling resources. Whether you want to train your existing teams or democratize data tools across your entire workforce, these programs will help you build an AI-ready organization:

SMB-Focused Programs: Explore our new SMB Learning Path or enroll in the Gemini Enterprise Agent Ready (GEAR) program for specialized training in agentic AI.
Google Skills for Organizations: Access our no-cost, on-demand learning platform featuring over 3,000 AI courses and hands-on labs created by experts at Google Cloud and Google DeepMind.
Get Certified: Ready to validate your team's expertise? This premium, cohort-based program offers instructor-led training, technical mentorship, and AI-infused skill badges designed to prepare your team for industry-recognized certifications.

By offering a full suite of SMB technology and training — from productivity in Workspace, to all our Ads services, and now powerful AI tools — Google is helping small and midsize firms thrive, no matter where the future takes us.

From AI potential to agentic reality: Driving the UK’s next chapter

Wed, 17 Jun 2026 08:00:00 +0000

The United Kingdom, and London in particular, continues to be one of the great hubs for AI development in Europe and the world. We’re home to Google DeepMind, of course, as well as significant AI unicorns — and Google Cloud customers — like Ineffable Intelligence, which is today announcing an important partnership with us.

A year ago, we joined you for the London Summit to showcase the vast potential of generative AI, including a major investment in upskilling the UK civil service. Today, as we welcome our partners once again to the historic vaults of Tobacco Dock, that potential has become an industrial-scale reality. In my conversations with leaders across both Whitehall and The City, the focus has moved from chatbots and media experiments to full-production execution. This is the moment of the agentic enterprise, where we shift from systems that simply chat with us to systems that can reason, plan, and execute multi-step workflows.

This transition is the cornerstone of the UK’s projected £400 billion economic boost from AI by 2030. At Google Cloud, we are the only provider offering the full integrated stack — custom silicon, frontier models, and planet-scale infrastructure — required to turn the Agentic Enterprise into a reality.

The new frontier of British enterprise and research

The banking sector is a key proving ground for this shift. And HSBC, one of the largest and most important financial institutions in the world, is showing the way. Today, we’re announcing a multi-year transformational partnership with HSBC to accelerate AI adoption across HSBC’s products and services globally. This new collaboration will further accelerate the shift towards AI-enabled ways of working across HSBC’s global operations.

HSBC will work with Google Cloud and Google DeepMind engineering teams to collaborate on new AI-powered tools and programmes, with access to Google’s latest agentic AI capabilities – including Gemini models and the Gemini Enterprise Agent Platform. The initial delivery focus on three areas: hyper‑personalised wealth management support, stronger financial crime risk management, and AI tools to enhance frontline/relationship manager client service

UK startups also continue to break new ground with technology, and AI in particular, as demonstrated by the work of frontier labs like Ineffable Intelligence. The company, which launched earlier this year, has chosen Google Cloud as its preferred cloud partner, utilizing Google’s full stack of AI-optimized hardware and tools to build and train Ineffable’s first generation of foundational models.

Led by David Silver, a former Google DeepMind researcher who was instrumental in the AlphaGo project, Ineffable Intelligence is taking a unique approach to AI development. The team are building systems that learn primarily through their own experience through reinforcement learning, instead of relying on the large-scale human-generated datasets behind language models. The ambition is to create a “superlearner” that develops knowledge through trial and error. This year, Ineffable Intelligence set a record for a European seed funding round of $1.1 billion, and now Ineffable Intelligence will support its training work by deploying one of the largest clusters of A5X, powered by the NVIDIA Vera Rubin NVL72 platform on Google Cloud, delivering massive computational scale.

To move from experimentation to true industrial production, businesses need more than just models; they need a roadmap. To help show them the way, we’re expanding our partnership with Deloitte, which will open a new AI Studio at its London campus. Developed in collaboration with Google Cloud, the studio will help British organisations move beyond AI experimentation to deploy autonomous, action-oriented AI systems at scale.

Deloitte is also committing to upskill 1,000 members of its UK AI and data workforce on Gemini Enterprise. This certification program will ensure that Deloitte’s AI and data engineers’ are equipped with the technical expertise to implement Google’s most advanced agentic architecture, providing UK clients with one of the largest pools of certified AI talent in the region.

Building a future-ready public sector

The blueprint for a modern digital government requires moving away from rigid legacy contracts toward agile, AI-driven public services. In collaboration with the Ministry of Housing, Communities and Local Government (MHCLG), the i.AI incubator, Google Deepmind, and Faculty, we are delivering tangible public sector reform and tools for reinvention that directly support the national goal to "get Britain building."

Agencies like MHCLG are already using a tool called Extract which was built using Google technology to help transform planning processes by reducing document processing times from two hours to just two minutes. Simultaneously, we are supporting trials of an AI planning tool — co-created with local planning authorities in Barnet, Dorset, and Camden — which aims to cut decision times for everyday applications by 50%. Furthermore, the Department for Transport (DfT) is utilizing Gemini to streamline public consultation analysis, a move projected to save £4 million annually.

Innovation on this scale also requires a secure, sovereign foundation. That is why Google Cloud is working to strengthen our UK data residency commitments, including measures like making Gemini 3.5 Flash, which features in-country AI processing, available by late June 2026 for sensitive sovereign use cases. We are giving British organizations the confidence to innovate within strict compliance boundaries.

To help keep businesses safe from the challenges posed by bad actors using AI and other digital threats, we also recently announced a comprehensive AI-powered cybersecurity platform — Google AI Threat Defense — which combines Wiz, Mandiant, Gemini & CodeMender to find, fix, and protect our customers from vulnerabilities.

Proven impact from the high street to public service

Autonomous agents are no longer a future prospect; they are delivering value across the UK economy today. Our work with THG Ingenuity, an ecommerce solutions provider, has delivered an 8x higher conversion rate via its AI Shopping Assistant. Starling is similarly empowering customers with "spending intelligence" tools for instant habit analysis around purchases and expenses. And Rightmove, has launched a beta version of an AI-powered conversational property search, built with Google’s Gemini models, enabling users to search for homes in their own words.

The breadth of this impact is visible across every sector: Kingfisher is pioneering retail-specific agentic applications; Openreach is driving field service optimization in telecommunications; andUnilever is using AI at scale across the entire value chain to drive growth and build desirable brands in the new era of consumer goods.

Meanwhile, VMO2 is streamlining complex data operations; Vodafone is executing a $1 billion partnership to redefine network performance; and WPP is integrating Gemini across creative workflows, whether that's generating high-fidelity campaign assets at speed and scale, powering AI agents, or training robotic camera operators.

Empowering the engine of growth for small to medium businesses and startups

The true measure of Britain’s AI success lies in its small and medium enterprises and startup ecosystem. Our AI Works research highlights a pivotal moment: AI has the potential to boost productivity for small and medium enterprises by 20% and unlock £198 billion in output for the UK economy. With 56% of smaller firms already seeking guidance, we have launched the AI Works for Britain upskilling initiative to ensure no business is left behind.

We also continue to foster the next generation of British unicorn startups through our ongoing partnership with Tech Nation at the London AI Hub. This sustained commitment ensures founders have the resources and community needed to scale, and this September, we will further this mission by hosting the Gemini Startup Forum: Cybersecurity in London to help startups build secure-by-design AI applications.

The Model Garden at Platform 37

Our belief in the UK’s potential is reflected in our physical footprint, too. We are continuing to invest in the UK's digital infrastructure to support growing demand: Our state-of-the-art data center in Waltham Cross launched in September 2025, a key part of our two-year, £5 billion investment to help power the UK's AI economy. And earlier this year, we opened our new office in London in Kings Cross, Platform 37, along with plans for The AI Exchange, a new public space dedicated to deepening understanding of AI.

Building on this momentum, we are excited to introduce The Model Garden at Platform 37, launching in the fourth quarter of 2026. This London-based hub is far more than a physical space; it serves as a strategic investment designed to fundamentally elevate how we engage with our most important customers. Blending the timeless aesthetics of a classic English garden with immersive, high-tech innovation — from living digital walls to a three-story atrium — The Model Garden acts as a physical marketplace for our best ideas.

The blueprint for the agentic enterprise

For UK businesses, civic leaders, and organizations to continue to lead in the AI moment, they must not only rethink the technology they use but also fundamental aspects of how we work. As we support thousands of organizations and millions of teams here and around the globe, we see three core strategies helping achieve success with AI:

Culture: We must reimagine our organizations for the future. True transformation means getting teams excited, enabled, and equipped to work with AI agents in completely new ways. It is about human-AI collaboration, not just automation.
Responsibility: We must build with safety and security in mind from day one. Protecting your users, your customers, and your brand is paramount. Our frontier models are built on a foundation of rigorous AI principles and secure-by-design infrastructure.
Sustainability: In an era of rising compute demands, we must scale in a way that is both financially viable and positive for our planet. At Google, we are committed to carbon-free energy 24/7, ensuring that the UK’s AI growth does not come at the cost of our climate goals.

Architecting the future together

Google Cloud is the primary partner for the UK’s agentic transition. We are moving beyond the hype of experimentation into the rigor of production. From the research labs of King's Cross to the diverse enterprises powering the high street, we are architecting a resilient, sovereign, and prosperous future for the United Kingdom.

Thank you to everyone who’s joining us in London — yesterday, today, and into the future. This year we’ve packaged up an exclusive on-demand experience, allowing you to stream the defining London Summit moments, available anywhere, anytime.

How Siemens "slices the elephant," advancing agentic workflows for industrial software development

Tue, 16 Jun 2026 07:00:00 +0000

For technology companies like Siemens, software is the nervous system of factories, energy grids, and transportation networks worldwide.

As a global leader in industrial AI, industrial software, and industrial automation, Siemens brings decades of domain expertise across factory and process automation, energy infrastructure, and intelligent transportation — expertise that no off-the-shelf AI solution can replicate. But innovation carries a heavy anchor: legacy code.

With codebases spanning hundreds of millions of lines developed for over more than a decade, Siemens faced a challenge that standard AI tools couldn't solve: understanding and modernizing this code and the applications which run on it. The scale and depth of industrial-grade software demand a fundamentally different approach. Existing coding assistants lacked the contextual depth required to navigate complex, multi-layered industrial codebases — a gap Siemens set out to close.

To solve this, Siemens and Google Cloud created Knowledge Fabric, an AI system for automating the software development lifecycle. It was built using knowledge graphs on Spanner Graph, the Google Agent Development Kit, Gemini API, Gemini Enterprise Agent Platform, Gemini CLI, and Anthropic Claude Code. In a pilot migrating existing frontiers to web-based interfaces, Knowledge Fabric reduced implementation effort, freeing engineers to focus on customer innovations while maintaining full system compatibility.

“By ingesting the entire software ecosystem into an intelligent agentic system equipped with custom knowledge graphs, we aren’t just helping developers optimize their development time; we are enabling autonomous agents to reason across the past to build the future,” said Franz Menzl, senior vice president, product creation excellence at Siemens. “This is about freeing engineers from repetitive work so they can focus on higher-value problem solving.”

The challenge: the complexity of industrial software

Modernizing large-scale industrial-grade software systems is often compared to rebuilding a jet while flying it. For Siemens, the challenge had four dimensions:

Scale: The repositories are massive — far exceeding the context windows of standard large language models.
Fragmentation: Critical knowledge was scattered across code, Jira tickets, Confluence pages, and scanned PDF manuals from the early 2000s.
Complexity: Tracing the link between a specific line of code and a functional requirement document from 10 years ago presented a challenge that no manual or conventional tooling approach could address efficiently. It’s a reality shared across the industry.
Responsibility: Systems must adhere to strict quality, compliance, and lifecycle requirements, often over 15 to 20 years of operation. AI‑generated outputs must therefore be explainable, traceable, and verifiable. Hallucinated or unvalidated changes are not merely inefficient but operationally unacceptable.

"We realized that standard RAG (retrieval-augmented generation) wasn't enough," said Agata Gołębiowska, technical lead, Google Cloud. "Code isn't just text; it has inherent structure. A class belongs to a file, which belongs to a module. Flattening that into a vector database meant losing the representation of relationships elements of the codebase."

The solution: A domain-aware Knowledge Fabric

To make this sprawling software environment navigable for AI-driven workflows, the teams built the Knowledge Fabric agent. This agent goes beyond keyword matching to “understand” the relationships between assets.

We use Spanner Graph to model the inherent structure of the codebase, applying the same rigor to documentation across formats. By mapping connections between these domains, we can link specific code snippets directly to requirements in a design document. Agents then traverse this graph, using tools to query the structure via Graph Query Language (GQL).

But GQL is only one piece. To enable semantic understanding, we generate embeddings for every node, using Spanner's Approximate Nearest Neighbors (ANN) algorithm to perform efficient vector search across the full codebase. Finally, we give agents full-text search capabilities, which can be combined with GQL to pinpoint nodes and edges with precision.

Combining these three methods lets an LLM agent answer complex queries, such as: "Which functions need to be updated if I change the logic in the Axis Control Panel?" The system traverses the graph — weighing keyword and semantic similarity — to identify dependencies, retrieve relevant documentation, and present a precise impact analysis.

This precise context is what lets a coding agent produce a valid, usable, and maintainable implementation.

"Slicing the elephant:" the agentic workflow

A key insight from the project was that AI agents struggle with massive, ambiguous tasks. To succeed, the team adopted a design pattern dubbed "slicing the elephant."

The system breaks a sweeping request like “refactor this module” into smaller, more manageable tasks, each handled by a specialized agent built with the Google Agent Development Kit (ADK):

Search agent: Acts as a deep-research specialist. It uses tools to explore the code graph and cross-reference findings with documentation in Agent Search.
User story agent: Interviews the product owner to gather requirements, then drafts detailed user stories with acceptance criteria linked to existing system contexts.
Architecture impact agent: Analyzes proposed changes against the graph to predict side effects before a single line of code is written.
Task breakdown agent: Consumes the analysis from the architecture impact agent and breaks the work into small, manageable tasks, each carrying all the context relevant to a specific change.
Coding agent: Implements the change described in a specific task. Reaching this step without context and prior analysis produces unusable code.

The system keeps a human in the loop at every step, which ensures reliable, production‑grade outcomes and keeps engineers focused on meaningful work rather than routine implementation.

"By slicing the elephant — breaking complex refactoring jobs into smaller, agent-led tasks — we observed a significant productivity increase," said Alexander Lomakin, project lead at Siemens. "We essentially gave the AI the roadmap it needed to navigate the complexity."

Pilot results: Faster, more efficient engineering

Developers saw results almost immediately.

Analyzing dependencies for a new feature once required senior engineers to spend several days navigating codebases and legacy documentation. With the Knowledge Fabric, the same work now takes far less time.

In a recent production pilot migrating legacy control panels to modern web‑based interfaces, the Knowledge Fabric reduced overall coding effort while preserving system integrity and industrial quality standards.

Engineers now spend more time creating customer value and less on repetitive work.

Get started

The Knowledge Fabric shows that generative AI can do more than write boilerplate code, it can also help teams modernize the legacy systems their businesses depend on most.

To learn more about building graph-based agents for your own legacy modernization:

Read about Spanner Graph.
Explore Agent Platform and find pre-built production-grade agents on Agent Garden
Check out the Agent Development Kit.
Read more on how Siemens is advancing industrial AI.

Cloud CISO Perspectives: The 4 lessons that guided AI Threat Defense

Mon, 15 Jun 2026 16:00:00 +0000

Welcome to the first Cloud CISO Perspectives for June 2026. Today, we introduce Chris Betz as the new CISO of Google Cloud. For his first Cloud CISO Perspectives, Chris shares four key lessons we learned about using AI to the defender’s advantage while building AI Threat Defense.

aside_block: <ListValue: [StructValue([('title', 'Get vital board insights with Google Cloud'), ('body', <wagtail.rich_text.RichText object at 0x7fa6550bf1f0>), ('btn_text', 'Visit the hub'), ('href', 'https://cloud.google.com/solutions/security/board-of-directors?utm_source=cgc-site&utm_medium=et&utm_campaign=FY26-Q2-GLOBAL-GCP39634-email-dl-dgcsm-CISOP-NL-177159&utm_content=-&utm_term=-'), ('image', <GAEImage: GCAT-replacement-logo-A>)])]>

Cloud CISO Perspectives: The 4 lessons that guided AI Threat Defense

By Chris Betz, CISO, Google Cloud

Chris Betz, CISO, Google Cloud

Just a year ago, it would take months or even years for a good application security team to find thousands of vulnerabilities. Today, a team equipped with multiple AI models can find the same number in hours — or even minutes.

AI is rewriting the rules of cybersecurity. It’s true that AI has boosted adversaries, introducing new threat actors, techniques, and surfaces to defend against, all operating with unprecedented scale, speed, and sophistication. AI-powered attackers are developing zero-day exploits by analyzing more than just source code: Configuration vulnerabilities, binaries, and firmware are all in their crosshairs.

However, AI has also created a significant advantage for defenders. Not only are these same capabilities in our hands, adding to our defense, but we have the added advantage of the full business context that adversaries lack. Software security, and especially vulnerability finding and fixing, is being revolutionized.

Security is changing rapidly, demanding that we all innovate in response. Here is how we are approaching this work today, and some of the lessons we learned along the way.

It’s clear that the AI benefits for security are rapidly evolving, and we can no longer rely on legacy, manual defenses. The new imperative for CISOs and business leaders is to transform vulnerability management by combating machine-speed threats with a defensive strategy that’s AI native, agentic, and open.

We’ve been preparing for this moment for years: From Project Naptime, an internal project to automate vulnerability hunting (so security researchers can take regular naps), to Big Sleep, our autonomous zero-day hunter, to CodeMender, our automated AI-patching agent, we’ve innovated to advance using AI to improve security for all.

Across our products and services, we’ve found that a unified approach helps us protect Google at Google scale. Based on this approach, we recently introduced AI Threat Defense as a pathway to achieve the threat-readiness transformation that you need to defend against AI threats with AI.

The framework is straightforward, and you’ll find that it’s ultimately about two key points:

Using rapidly-advancing AI to protect ourselves.
Shifting the way we develop from the ground up.

Security is changing rapidly, demanding that we all innovate in response. Here is how we are approaching this work today, and some of the lessons we learned along the way.

Four key lessons

Our work is built on a four-step framework, structured directly on what we learned:

Prepare: How Google started the journey — hardening our foundation and operationalizing the framework.
Scan and prioritize: How we identified vulnerabilities — conduct deep-dive analysis and posture validation.
Remediate: What we learned from remediation — implement workflows to autonomously verify and patch vulnerabilities quickly.
Monitor: How we evolved monitoring with AI agents — transition to continuous detection and active response playbooks.

1. Prepare: A modern enterprise runs on an enormous amount of software, and at Google that amount is even greater. We needed focus in order to move at speed, so our first lesson was to reduce our attack surface. That let us narrow our focus, reduce complexity, and use insights we have on our software supply chain and dependencies to prioritize and protect our external interfaces.

Second, we invested in the operational framework supporting the vulnerability work. Early experimentation quickly showed us how valuable a scaling framework is that applies our knowledge of the environment, protects and allocates resources for scanning, and allows new capabilities to be iterated on and used by multiple teams. The amplifying power of good information, code access, dependency graphs, token budgets, and infrastructure are key friction reducers.

Third, we planned engineering work alongside security work: Your engineering partners are critical, especially for aligning with your resiliency and deployment processes.

How Google Cloud CISO Chris Betz Uses LLMs to Defend Billions of Users from Vulnerablities

Key lessons include:

Tagging components with the model, harness, and issues found when scanning.
Allocating hardware and token budgets for finding, developing fixes, build and test.
Managing change volume (and engineer hours) while simultaneously focusing on more, smaller updates, where possible, with good rollout plans to de-risk the change.

2. Scan and prioritize: We continuously scan our code across products — Search, Ads, Android, Chrome, and Google Cloud — managing tens of thousands of packages.

First, we kicked off scanning and centrally tracked our progress, integrating the same tools into our pipelines. We learned early on that the best scanning results come from a combination of an expert in the specific product plus the harness plus the AI model. The combination is crucial, because results will be markedly different without all three.

It’s worth noting that if you can only pick two, we recommend expertise and harness. A less capable model with a good harness and good expert is more powerful than the best model without a good harness or good experts. We also advise using more than one model.

It’s important to track and iterate the data. Since the technology is evolving fast, your data is critical to revise and refine your processes.

Second, look carefully at your software supply chain, and engage your key suppliers. Reachability remains a key criteria for fixes, as does streamlining and simplifying the areas you work on.

Third, because there are so many vulnerabilities that can show up, it’s important to have the right methodology to prioritize them. Normally, when you’re rolling out a change you prioritize the smallest blast radius to make incremental change. Here, we recommend flipping that model: Begin with foundational code with the biggest blast radius to tackle the hardest problems first.

AI models can do a good job of developing proof-of-concepts to rapidly test accuracy. Harness and models play a significant role in reducing false positive rate. Adapting your harness to do validation and using a different agent or model to validate results are both very valuable.

Another key to AI-powered triage is to use your harness and tools to state vulnerability confidence as well as severity. Of course, developing a patch is only part of the problem.

3. Remediate: Fixing vulnerabilities at Google scale required a fundamental shift in strategy. We developed a new approach centered on three lessons.

First, how you roll out patches matters. We adopted a risk-based approach that prioritized code reachable from the outside and had the largest blast radius, such as critical applications like BoringSSL and gVisor. We also learned that providing the model with context was the key to faster, more trusted remediation.

Second, we learned you cannot fix what you cannot track. To manage remediation at scale, we built a central system to track every vulnerability, from discovery to resolution, with every finding labeled in a central repository. This single source of truth allowed us to enforce service-level objectives (SLOs) for patching, and enabled us to deploy constant autonomous patching with human review. Coupled with robust roll-back capabilities, our teams got better at fixing things quickly and safely.

Finally, we learned to build resilience directly into the system. The ultimate goal was to create an inherently-resilient system that can also patch vulnerabilities, not the other way around. We don't just fix the code; we harden the entire system around it.

These changes helped us rethink our approach to securing open-source software with a three-R’s strategy: Refresh, remove, and rewrite.

First, we refresh what is foundational — finding and fixing vulnerabilities in the code. This is about being good network citizens and protecting the core.
Second, we remove what is peripheral. We are removing dependencies and replacing them with custom code. This is about both efficiency and reducing the attack surface, moving from a broad base of trust to a narrow, controlled one.
Third, we rewrite what is critical. For everything in between, we are transitioning legacy logic and critical capabilities into modern, memory-safe languages using AI to automate the transition to eliminate entire classes of vulnerabilities from that software.

This evolution is a deliberate approach to reduce complexity, shrinking the attack surface, and building a more resilient, autonomous, and secure-by-design foundation for everything we do.

4. Monitor: Our work doesn’t stop there, and neither should yours. The security landscape is always changing, and the monitor phase is where our approach comes alive by creating a perpetual feedback loop to ensure we stay secure — and get stronger over time.

We had three key lessons in this phase. First, security demands a constant feedback loop. We created a feedback loop to monitor the entire ecosystem for two things: system strain and vulnerability hotspots.

Second, we invested in tracking our long-term remediation health. You can only improve what you measure. We built a comprehensive asset inventory to track our overall security posture and the completeness of our remediation efforts. Here’s where we hold ourselves accountable to product-level SLOs for vulnerability management.

This system allows us to deploy rolling patches that can update even our data center hardware continuously and use AI agents to verify patch efficacy at a scale no human team could manage.

Third, we planned for the future by using AI agents for both coding and monitoring. You have to assume that at some point, the attackers' models will become more advanced. We need to evolve our operating model and build for that reality.

We use AI agents to automate and standardize our response playbooks, enabling instantaneous containment when an issue is found. We move beyond just finding bugs by feeding key libraries into Gemini to improve its pattern recognition, creating security-aware coding agents. Meanwhile, our AI-assisted red teamers are continuously stress-testing our core infrastructure, ensuring our defenses are always evolving.

The outcome of this constant monitoring is a living, measured program that we can trust.

This is how we protect billions of users every day, and it provides a framework that any team can use to build a defense that learns, adapts, and hardens itself against the threats of tomorrow.

To learn more about AI Threat Defense, you can watch our recent Security Talks online event.

aside_block: <ListValue: [StructValue([('title', 'Learn something new'), ('body', <wagtail.rich_text.RichText object at 0x7fa6550bf250>), ('btn_text', 'Watch now'), ('href', 'https://www.youtube.com/watch?v=blh0hhHJ4pI'), ('image', <GAEImage: Cloud-CISO-Perspectives-logo-A>)])]>

In case you missed it

Here are the latest updates, products, services, and resources from our security teams so far this month:

Detecting and containing AI-powered threats with Google Security Operations agents: Learn how Google Security Operations works in concert with AI Threat Defense to monitor, detect, and respond to threats, particularly from code you do not own or can not patch. Read more.
How to stop AI voice clones from bypassing your security perimeter: The traditional, relatively stable network perimeter has been replaced by one far more malleable: Identity, driven by vishing attacks. Here’s how to defend against them. Read more.
5 lessons from red teaming AI applications: Distilled from Mandiant’s hands-on red team experiences, check out our clear, concise guidance to help customers securely develop and deploy AI apps. Read more.
Introducing Wiz Cloud Cost: Powering cost management and optimization with context: Wiz unifies cloud and AI cost visibility to help teams eliminate waste and improve spend efficiency across their AWS, Azure, and Google Cloud environments. Read more.
Bringing AI agents to Chrome Enterprise security management: We're launching an open-source model context protocol (MCP) server that connects AI agents directly to Chrome Enterprise APIs, helping IT and security teams manage browser security more efficiently. Read more.
How Google Does It: An inside look at cybersecurity: Learn how Google approaches some of today's most pressing security topics, challenges and concerns, straight from Google experts. View the collection.

Please visit the Google Cloud blog for more security stories published this month.

aside_block: <ListValue: [StructValue([('title', 'Join the Google Cloud CISO Community'), ('body', <wagtail.rich_text.RichText object at 0x7fa6550bf2b0>), ('btn_text', 'Learn more'), ('href', 'https://rsvp.withgoogle.com/events/google-cloud-ciso-community-interest-form-2026?utm_source=cgc-blog&utm_medium=blog&utm_campaign=FY25-Q1-global-GCP30328-physicalevent-er-dgcsm-parent-CISO-community-2025&utm_content=cisop_&utm_term=-'), ('image', <GAEImage: GCAT-replacement-logo-A>)])]>

Threat Intelligence news

Seeking counsel: Ongoing targeted campaign against U.S. law firms: Mandiant Consulting details a financially-motivated data theft extortion campaign executed by the threat cluster UNC3753, highlighting tactics like physical office targeting, and provides actionable recommendations to safeguard endpoints and infrastructure. Read more.
Welcome to BlackFile: Inside a vishing extortion operation: Google Threat Intelligence Group (GTIG) has continued to track an expansive extortion campaign by UNC6671, a threat actor operating under the "BlackFile" brand, that targets organizations via sophisticated voice phishing (vishing) and single sign-on (SSO) compromise. Read more.
2 PhaaS 2 Furious: The evolution of Chinese-language phishing services: While Russian-speaking threat actors have historically dominated the phishing-as-a-service (PhaaS) landscape, a rival ecosystem is rapidly growing within the Chinese-language underground. Within this ecosystem, GTIG has observed a fundamental move away from static password harvesting towards real-time interception and tokenization. Read more.

Please visit the Google Cloud blog for more threat intelligence stories published this month.

Now hear this: Podcasts from Google Cloud

Cloud Security Podcast: Deceiving adversaries at scale: Kevin Conley from Riot Games discusses how modern organizations can use deception technology to gain a home-field advantage against adversaries by proactively monitoring their environments. Listen here.
Cloud Security Podcast: Hyperscaling cloud security with Wiz: Yinon Costica, co-founder and VP of product, Wiz, discusses how the company used a product-led approach and a unique security graph model to scale rapidly within the competitive cloud security market. Listen here.
Behind the Binary: When AI features create zero-click exploits: Google Project Zero’s Seth Jenkins joins the podcast to dissect a full two-bug, zero-click exploitation chain targeting the Pixel 9. Listen here.

To have our Cloud CISO Perspectives post delivered twice a month to your inbox, sign up for our newsletter. We’ll be back in a few weeks with more security-related updates from Google Cloud.

Introducing the Open Knowledge Format

Fri, 12 Jun 2026 13:00:00 +0000

As foundation models continue to improve, the lack of relevant context often limits what they can do, especially as they are used to build agentic systems. While these models can help you write code, summarize documents, or analyze a dataset, they still need the right information to produce accurate and actionable results.

That’s why today, we’re introducing the Open Knowledge Format (OKF), an open specification that formalizes the LLM-wiki pattern into a portable, interoperable format. This is a vendor-neutral, agent- and human-friendly standard for representing the metadata, context, and curated knowledge that modern AI systems need.

As published, OKF v0.1 represents knowledge as a directory of markdown files with YAML frontmatter, with a small set of agreed-upon conventions that let wikis written by different producers be consumed by different agents without translation.

That's it. No complex compression scheme, no new runtime, no required SDK. A bundle of OKF documents is:

Just markdown — readable in any editor, renderable on GitHub, indexable by any search tool
Just files — shippable as a tarball, hostable in any git repo, mountable on any filesystem
Just YAML frontmatter — for the small set of structured fields that need to be queryable: type, title, description, resource, tags, and timestamp

If you've used Obsidian, Notion, Hugo, or any of the LLM wiki patterns that have emerged over the past year, the shape will feel familiar. OKF formalizes the small set of conventions needed to make these patterns interoperable.

Let’s take a look at the problem that OKF can solve for your organization, how it works, how to get started with it, and what’s next.

A fragmented context landscape

In most organizations, the information that foundation models use is overwhelmingly internal knowledge: the schema of a table, your business’ meaning of a metric, the runbook for an incident, the join paths between two systems, the deprecation notice for an old API, etc.

Today, these atoms of knowledge live in a variety of highly fragmented systems:

Metadata catalogs with their own APIs
Wikis, third-party systems, or in shared drives
Code comments, docstrings, or notebook cells
The heads of a few senior engineers

When an AI agent needs to answer "How do I compute weekly active users from our event stream?" it has to assemble the answer from these scattered, mutually incompatible surfaces. Every vendor offers its own catalog, its own SDK, its own knowledge-graph schema, and none of the knowledge is easily portable across products or organizations.

The result: Every agent builder is solving the same context-assembly problem from scratch, every catalog vendor is reinventing the same data models, and the knowledge itself is locked behind whichever surface created it.

Knowledge as a living wiki

Developer teams are changing how they build AI agents. Instead of using models to search the same documents for the same facts over and over, you can give your agents a shared markdown library that grows more useful over time. This lets your agents take on the drudgery of reading and updating their own files, while your team curates the content and manages it like code.

Andrej Karpathy, the prominent AI researcher and educator, articulates this idea most crisply in his LLM Wiki gist. "LLMs don't get bored, don't forget to update a cross-reference, and can touch 15 files in one pass," he writes. The bookkeeping that causes humans to abandon personal wikis is exactly what LLMs are good at.

Similar knowledge-as-Wiki pattern keeps reappearing under different names: Obsidian vaults wired to coding agents, the AGENTS.md / CLAUDE.md family of convention files, repos full of index.md and log.md artifacts that agents consult before doing real work, and "metadata as code" repositories inside data teams.

The pattern is compelling and powerful, but each instance is bespoke. Karpathy's wiki and your team's wiki and a vendor's catalog export may all look alike (markdown, frontmatter, cross-links), but none of them are intentionally designed to cooperate. There is no agreed-upon answer to what fields every document should carry, or what filenames mean what. As a result, the knowledge encoded in wikis remains siloed within the original teams, leading to redundant effort whenever a new agent is built.

What's missing is a format, not another service

The answer to this problem isn’t another knowledge service. You need a format, a way to represent knowledge that:

Anyone can produce, without an SDK
Anyone can consume, without an integration
Survives moving between systems, organizations, and tools
Lives in version control alongside the code it describes
Is readable by humans and parseable by agents: the same file, no translation layer

By design, OKF is that format.

How OKF works: The design in one screen

An OKF bundle is a directory of markdown files representing concepts: anything you want to capture, including tables, datasets, metrics, playbooks, runbooks, and APIs. Each concept is one file. The file path is the concept's identity:

code_block: <ListValue: [StructValue([('code', 'sales/\r\n├── index.md\r\n├── datasets/\r\n│ ├── index.md\r\n│ └── orders_db.md\r\n├── tables/\r\n│ ├── index.md\r\n│ ├── orders.md\r\n│ └── customers.md\r\n└── metrics/\r\n│ ├── index.md\r\n └── weekly_active_users.md'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa6550fa1f0>)])]>

Each concept document has a small block of YAML front matter for structured fields and a markdown body for everything else:

code_block: <ListValue: [StructValue([('code', '---\r\ntype: BigQuery Table\r\ntitle: Orders\r\ndescription: One row per completed customer order.\r\nresource: https://console.cloud.google.com/bigquery?p=acme&d=sales&t=orders\r\ntags: [sales, revenue]\r\ntimestamp: 2026-05-28T14:30:00Z\r\n---\r\n\r\n# Schema\r\n\r\n| Column | Type | Description |\r\n|---------------|-----------|------------------------------------------|\r\n| `order_id` | STRING | Globally unique order identifier. |\r\n| `customer_id` | STRING | FK to [customers](/tables/customers.md). |\r\n\r\n# Joins\r\n\r\nJoined with [customers](/tables/customers.md) on `customer_id`.'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa6550fa250>)])]>

Concepts link to each other with normal markdown links, turning the directory into a graph of relationships that is richer than the parent/child links implied by the file system. Bundles can optionally include index.md files (for progressive disclosure as agents navigate the hierarchy) and log.md files (for chronological history of changes).

The full v0.1 specification (including conformance criteria, cross-linking rules, and the small number of reserved filenames) fits on a single page.

Three principles behind the design

1. Minimally opinionated. OKF requires exactly one thing of every concept: a type field. Everything else (e.g., what types exist, what other fields to include, what sections the body has) is left to the producer. The spec defines the interoperability surface, not the content model.

2. Producer/consumer independence. OKF cleanly separates who writes the knowledge from who consumes it. A bundle hand-authored by a human can be consumed by an AI agent. A bundle generated by a metadata export pipeline can be browsed in a visualizer. A bundle synthesized by one LLM can be queried by another. The format is the contract; the tooling at each end is independently swappable.

3. Format, not platform. OKF is not tied to any specific cloud, database, model provider, or agent framework. It will never require a proprietary account or SDK to read, write, or serve. We're publishing it as an open standard because the value of a knowledge format comes from how many parties speak it, not from who owns it.

What we're shipping with the spec

To make the format concrete, we're publishing reference implementations at both the producer and consumer ends:

An enrichment agent that walks a BigQuery dataset, drafts an OKF concept document for every table and view, then runs a second LLM pass that crawls authoritative documentation and enriches each concept with citations, schemas, and join paths.
A static HTML visualizer that turns any OKF bundle into an interactive graph view in a single self-contained file; no backend, no install on the viewing side, no data leaves the page.
Three ready-to-browse sample bundles: GA4 e-commerce, Stack Overflow, and Bitcoin public datasets, produced by the reference agent and committed to the repo as living examples of conformant OKF.

These are proofs of concept, deliberately. The agent demonstrates one way to produce OKF; nothing about the format requires a specific agent framework or LLM. The visualizer demonstrates one way to consume it; nothing about the format requires HTML or a graph view. We expect (and want!) the ecosystem of producers and consumers to grow far beyond what we've shipped.

Where we go from here

OKF v0.1 is a starting point, not a finished standard. The format will evolve as more producers and consumers emerge and as we collectively learn what knowledge representations agents actually need in practice.

We're publishing in the open from day one because that's the only way a knowledge format earns its name, whether you're building a knowledge catalog, an enrichment pipeline, a wiki tailored to AI agents, or anything in the AI knowledge domain.

From here, we encourage you to:

Read the spec (it's short!)
Write a producer for your source system, your database, your documentation site
Write a consumer: a viewer, a search index, an agent that reasons over bundles
Try the reference implementation against your own data
File issues, send PRs, or propose extensions: The spec is versioned and explicitly designed for backward-compatible growth

The repo, the spec, and the sample bundles are available in GitHub. We have also updated Google Cloud’s Knowledge Catalog to be able to ingest Open Knowledge Format and serve it to our agents. You can find the relevant code and examples here.

The format itself is the contribution. The tools we've shipped exist to make it real, and to lower the cost of trying it out. Whatever shape your knowledge takes today, OKF is designed to be the lingua franca it can be exchanged for tomorrow.

^{Published by the Google Cloud Data Cloud team. Open Knowledge Format is an open specification; contributions, alternative implementations, and adoption beyond Google products are all explicitly welcomed.}

^{In addition to the authors, this work came together thanks to key ideas from many others at Google, and we thank them for their contributions.}

Powering the next era of Confidential AI

Thu, 11 Jun 2026 19:30:00 +0000

At Google Cloud, we’re committed to providing the most advanced, secure, and private infrastructure for the most demanding AI workloads, and partnering with a broad and diverse range of organizations to help them meet their AI workload needs.

We are thrilled to collaborate with Apple on its expanded Private Cloud Compute (PCC) systems announced this week at WWDC 2026. Working closely together, Apple and Google have built a serving platform on Google Cloud that meets the rigorous security, confidentiality, and transparency goals that Apple has for PCC. This achievement is a testament to the strong collaboration between our teams, as well as with Intel and NVIDIA.

Our commitment to privacy with Confidential Computing

Our collaboration with Apple is built on a foundation of deep commitment to privacy that leverages Google Cloud's security and privacy technologies. At the heart of this collaboration is our Confidential Computing portfolio and our Titanium security architecture.

Titanium architecture, featuring our custom-designed Titan chip, provides a hardware root of trust that underpins the security and integrity of Google's infrastructure and services. Confidential Computing builds on this secure foundation by helping ensure data is protected throughout the lifecycle, encrypted at rest, in transit, and crucially in use within hardware-based Trusted Execution Environments (TEEs).

By protecting data in use, Confidential Computing becomes a fundamental and foundational element for building trust in AI systems, providing verifiable integrity and isolation for sensitive workloads. Confidential Computing helps prevent unauthorized access because data remains encrypted and isolated.

Enabling Apple Private Cloud Compute on Google Cloud

We are proud to collaborate with Apple to extend the privacy and security properties of PCC infrastructure to Google Cloud. Our platform supports Apple’s PCC privacy commitments with a layered security approach built upon Google Cloud’s infrastructure, including:

Google Cloud Confidential Computing: Our core Confidential Computing platform provides the hardware-based TEEs necessary for PCC. By leveraging Intel TDX (Trust Domain Extensions) and NVIDIA Confidential Computing, we provide hardware-based isolation for virtual machines, designed to create a highly secure and private environment where workloads can run with cryptographic assurances.
Google Titanium security architecture and Titan chip: Google Titan chips are a key component in powering security and transparency posture for PCC infrastructure on Google Cloud. Deployed across our fleet, Titan establishes a strong hardware root of trust, helping to ensure the integrity of the boot process and the hardware platform itself.
Intel TDX and NVIDIA Confidential Computing: Google Cloud leverages the security features on Intel CPUs and NVIDIA Blackwell GPUs to protect data-in-use during high-performance AI inference, helping ensure that the entire compute path – from CPU to GPU – is protected.
Open-source transparency: With our commitment to verifiable security, Apple and Google have collaborated in engineering an open-source host stack specifically to support PCC's transparency, enabling independent inspection and verification of the system's security properties.

Together, these technologies help ensure that Apple PCC on Google Cloud meets requirements with enforceable protections, no privileged runtime access, and verifiable transparency.

Building the future of private AI infrastructure

Our collaboration with Apple represents a significant milestone in further strengthening a secure cloud for AI by building on technologies and standards from Apple, Google Cloud, Intel, and NVIDIA. By ensuring that every layer of the stack — both hardware and software — contributes to a verifiable and secure system, we’ve created an advanced platform that is designed to uphold the stringent standards of user privacy and data security that PCC architecture demands.

The advancements built through this collaboration will benefit all Google Cloud customers. We are committed to continuous improvement and offering more transparent, secure, resilient platforms for all types of workloads, especially those handling AI and sensitive data.

You can learn more about Confidential Computing here.

Claude Fable 5: Available on Google Cloud

Tue, 09 Jun 2026 18:00:00 +0000

Claude Fable 5, Anthropic’s latest frontier model, is now generally available on Google Cloud. This launch is the latest proof point of our ongoing commitment to bring the industry's latest models straight to our Agent Platform.

Claude Fable 5 brings the best of Anthropic model capabilities to all customers, with strong safeguards designed to make it safe for general use. Designed for complex, multi-step reasoning, Claude Fable 5 is good for demanding tasks like advanced software development, long-horizon agents, and deep multimodal document analysis. For more information about this release, visit Anthropic’s blog.

Build with Claude Fable 5 and other models from Anthropic — including Claude Opus 4.8 and Claude Sonnet 4.6 — today on Agent Platform.

How to unlock true ROI in software development – a deep dive into the latest DORA research

Tue, 09 Jun 2026 16:00:00 +0000

How do you prove the business value of generative AI to your teams?

Technology and finance leaders need to show the clear business value of AI projects to secure ongoing funding. While measuring return on investment (ROI) is a key part of validating your technical strategy, long-term success ultimately depends on building the organizational systems and culture needed to make AI work.

To help you evaluate the costs and business benefits of AI, we recently shared the DORA: ROI of AI-assisted software development report. This research offers a practical approach to help your team work through early adoption challenges, align engineering plans, and drive business growth.

Here are the key findings from the report, and how you can use them to support your overall technology strategy.

Insight #1: Navigating the J-curve of AI value realization

It is important to be realistic about how quickly you will see a return on your AI investments. While AI can act as a powerful amplifier for software engineering, the path to financial value is rarely a straight line. Most organizations will instead encounter a J-curve: a temporary productivity dip and period of instability associated with early adoption.

This temporary drop is a normal part of adopting new technology, rather than a sign of a failing strategy. The report points to three main reasons why this happens:

The learning curve: Teams require dedicated time away from regular feature delivery to adapt their daily workflows and master advanced techniques, evolving from simple prompting to building systems based on context and intent.
The verification tax: Because AI dramatically increases the sheer volume of code produced, developers must invest extra time rigorously reviewing generated outputs to ensure trustworthiness, prevent hallucinations, and meet internal architectural standards.
Pipeline adaptation: As individual developers generate code significantly faster, downstream processes like testing and change approvals often become bottlenecks and must be actively scaled to handle the increased throughput.

Budgeting for this initial learning phase is key to making the transition work. By anticipating this temporary drop in productivity, you can confidently keep your AI projects moving forward, knowing that these early challenges are an investment in your team's long-term speed.

The J-Curve of AI value realization

Insight #2: Understand the market divide on AI returns

DORA’s state of AI-assisted software development report shows that 90% of DORA survey respondents report using AI at work. Despite nearly universal adoption, actual financial impacts vary across organizations. Across the market, some companies see clear value from their engineering investments, while others struggle with unexpected costs.

When a project falls short, it’s often because the team lacks the organizational support to make it work. To get the returns you expect, you need to prepare your workflows and teams to adopt the new technology.

Insight #3: Calculating your AI ROI is essential

Building a realistic financial model for AI starts with looking at where it actually adds value. Across the software development lifecycle, AI can help your team reduce costs, boost productivity, improve security, and deliver a better experience for both developers and users.

To assist in modeling this for your organization, you can use this interactive ROI calculator.

This tool helps you explicitly forecast both the visible expenses and the hidden realities of AI adoption.
You can explore the mechanics, adjust the assumptions to match your reality, and build your own estimate.

The value model—from adoption to ROI

Get started

Download the full report: Explore the complete framework to quantify your AI investments, navigate the J-Curve, and map your AI investment roadmap.
Try out the interactive ROI calculator: Visit https://dora.dev/ai/roi/calculator to estimate your organization's potential returns and build a defensible business case.
Watch this Cloud OnAir webinar: From cost center to value engine: Building your business case for AI-assisted development.

Report: GKE Inference Gateway delivers up to 92% faster AI responses

Tue, 09 Jun 2026 16:00:00 +0000

As generative AI moves from experimental pilots to massive production environments, the efficiency of your infrastructure becomes the ultimate differentiator. One way to get the most out of it and minimize costly accelerator idle time is to leverage the Google Kubernetes Engine (GKE) Inference Gateway, which intelligently routes generative AI workloads based on real-time model server metrics.

Instead of relying on traditional, naive round-robin load balancing — which frequently triggers expensive accelerator recomputation and spikes user latency — this native extension of the GKE Gateway utilizes advanced capabilities like prefix caching and model-aware routing. By ensuring requests land on the exact accelerator that is primed to process them right away, GKE transforms how you can serve your large language models (LLMs), with excellent hardware utilization and ultra-fast response times.

In fact, according to an independent benchmark report, GKE Inference Gateway outperforms the next leading managed Kubernetes service with 15.7% higher throughput, 92.8% shorter wait times, and 62.6% lower inter-token latency. This performance takes LLM-based applications from sluggish and expensive to fast and production-grade.

That performance tracks with Snap’s experience using GKE Inference Gateway.

“At Snap, we are integrating llm-d into our production AI infrastructure to facilitate high-performance inference at scale. By employing prefix-cache-aware routing, we have achieved prefix cache hit rates ranging up to 75-80%. We appreciate the open-source nature of llm-d, as it enables seamless integration with our Envoy-based Service Mesh.” - Vinay Kola, Senior Manager, Software Engineering, Snap Inc.

In this blog, we take a closer look at GKE Inference Gateway’s prefix caching, complete with examples. We also provide more details about its benchmark results. Let’s jump in.

The secret to low-latency AI: Prefix caching

Prefix caching optimizes LLM performance by storing the KV cache (activation states) of long, repetitive prompt prefixes. When consecutive user requests share the same system instructions, context, or documentation, the model entirely skips reprocessing those tokens. GKE Inference Gateway reads incoming request prefixes and matches them to the specific pods that already hold that data in memory. This eliminates the "thinking" tax on your GPUs and TPUs, turning heavy reasoning loops into near-instant answers.

Use case 1: Documentation and codebase Q&A with retrieval-augmented generation (RAG)

When querying massive enterprise repositories, you can ground your LLMs’ responses without any added latency by pinning entire documentation sets as static cached prefixes, using RAG.

Instead of forcing an LLM to re-read thousands of lines of API references or corporate wikis for every single user question, GKE Inference Gateway routes the query to a pod that already has that specific context warmed up in its KV cache. The LLM only has to compute the user's brief, dynamic question, completely bypassing expensive document re-evaluation.

code_block: <ListValue: [StructValue([('code', '[STATIC PREFIX - STAYS IN CACHE] You are an expert AI assistant specializing in technical documentation. Below is the complete API documentation for our software platform. Use this context to answer the user\'s questions accurately. If the answer cannot be found in the documentation, say "I cannot find that in the provided context." \r\n\r\n<documentation> [10,000+ words of API reference documentation, endpoints, error codes, etc.] </documentation> \r\n\r\n[DYNAMIC SUFFIX - CHANGES PER REQUEST] User Question: How do I handle a 429 rate limit error using the Python SDK?'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654e2aa90>)])]>

Use case 2: Multi-turn chat

You can also use prefix caching to maintain customer service interactions across thousands of simultaneous sessions without compounding compute costs. You can do so by caching permanent system personas and core business rules directly on the LLM server.

In enterprise chat architectures, the base system prompt and reference tables remain completely identical across millions of customer interactions. GKE Inference Gateway handles these multi-turn conversations using context-aware routing to bypass repetitive token processing, so that your chatbot stays ultra-responsive even under peak traffic.

code_block: <ListValue: [StructValue([('code', '[STATIC PREFIX - STAYS IN CACHE] \r\n-System Persona: You are "FinBot", a helpful, empathetic, and compliant virtual assistant for ABC Banking Solutions. You must strictly adhere to the following rules: 1. Never provide concrete investment advice. 2. Always verify if the user is asking about checking or savings. 3. Keep your answers under 3 sentences. 4. If a user is angry, offer to connect them to a human manager. \r\n\r\nHere is the current interest rate table for May 2026: \r\n- Savings: 4.2% APR \r\n- Checking: 0.5% APR \r\n- CD (12-month): 5.1% APR \r\n\r\n[DYNAMIC SUFFIX - CHANGES PER REQUEST] User: Hi, I\'m trying to figure out how much I\'d make if I locked away $10,000 for a year?'), ('language', ''), ('caption', <wagtail.rich_text.RichText object at 0x7fa654e2ae80>)])]>

GKE outperforms alternative managed Kubernetes solutions

To validate these architectural advantages, Principled Technologies recently released an independent benchmark report comparing GKE (equipped with the GKE Inference Gateway) against a standard third-party managed Kubernetes service utilizing conventional round-robin HTTP load balancing.

Tested on a Llama 3.1 8B Instruct shared prefix workload using identical hardware (eight NVIDIA A100 40GB GPUs) the results reveal a massive performance gap between the two Kubernetes services. GKE didn't just win; it completely redefined inference efficiency across three critical metrics:

Higher throughput: 15.7% more tokens processed per second, enabling higher request capacity or reduced hardware needs for the same workload
Much faster time to first token (TTFT): 92.8% shorter wait times, producing dramatically quicker perceived response starts for interactive scenarios
Lower inter-token latency (ITL): 62.6% reduction, resulting in smoother and faster token streaming after the first token

Figure 3: Mean latency (normalized time per output token) of GKE with GKE Inference Gateway and third-party managed Kubernetes service on the Llama 3.1-8B Instruct LLM on the Shared prefix use case. Both solutions used the same hardware. Source: Principled Technologies

	GKE	3rd party ManagedKubernetes Service	GKE Advantage
Mean outputtoken throughput	7,169.21 outputtokens per second	6,042.05 outputtokens per second	15.7% more outputtoken throughput
Mean time tofirst token (TTFT)	188.36 ms	2624.73 ms	92.8% less TTFT
Mean inter-tokenlatency (ITL)	30.20 ms	81.03 ms	62.6% lower ITL

Figure 4: GKE with GKE Inference Gateway delivered superior AI inference compared to a third-party managed Kubernetes service using standard HTTP LB.

Ready to accelerate your gen AI inference workloads?

Whether you’re deploying inference workloads such as real-time customer support agents, dynamic coding assistants, or sub-second fraud detection models, infrastructure latency dictates your user experience. By ensuring shared prompt prefixes hit the active cache nearly 100% of the time, GKE Inference Gateway transforms your LLMs from sluggish, expensive reasoning engines into rapid, capital-efficient, production-grade powerhouses.

Ready to explore the performance advantage that GKE Inference Gateway can bring to your gen AI workloads? Access the full benchmark report here and watch this explainer video to learn more.

^{A special thanks to Dan Sullivan, Senior Performance Architect, Principled Technologies.}

Detecting and containing AI-powered threats with Google Security Operations agents

Tue, 09 Jun 2026 09:00:00 +0000

To defend against the growing range of AI-accelerated threat actors, organizations need to be able to respond faster to outpace the adversary.

Recently, we announced Google AI Threat Defense, an automated security system designed to help you continuously monitor for and stop AI-powered threats before they can impact your business. Based on Google’s own approach to today’s threats and vulnerability management, it’s centered on a four-step framework: Prepare, scan and prioritize, remediate, and monitor.

Today, we’re sharing more details on how Google Security Operations works in concert with AI Threat Defense to monitor, detect, and respond to threats, particularly from code you do not own or can not patch. The remediation gap represents a critical vulnerability.

According to M-Trends 2026, the exploitation of vulnerabilities has become the most common initial infection vector. Notably, the report also indicates that the mean time to exploit has dropped to an estimated minus seven days, meaning exploitation frequently occurs even before a patch is officially released. Google Security Operations delivers vital operational fabric to autonomously contain active attacks across your entire environment.

Google Security Operations supports AI Threat Defense to monitor, detect, and respond to threats.

Engineered around a comprehensive approach that uses compensating controls with proactive security to strengthen operational resilience, Google Security Operations is built on a strategic, three-part approach to cross-environment visibility across your entire attack surface:

Continuous and autonomous coverage analysis and detection generation
Autonomous investigation, containment, and response
Retroactive hunting

Designed to help you see and respond to threats faster than ever before, we deliver these capabilities at machine-scale and machine-speed. Together with Google AI Threat Defense, we’re able to provide the autonomous platform you need to outpace AI-driven attacks.

1. Continuous and autonomous coverage analysis and detection generation

While proactive defense can identify vulnerabilities before they can be exploited, there will be applications that you can not patch, as well as potential gaps in the time it takes to remediate vulnerabilities.

The 2026 Verizon Data Breach Investigations Report underscores the magnitude of this challenge. In a study encompassing over 13,000 organizations, only 26% of vulnerabilities identified on the CISA Known Exploited Vulnerabilities (KEV) list had been fully remediated. Moreover, the median duration required to achieve full patching after detection stands at 43 days. Clearly, you still need continuous monitoring to detect threats in your environments.

Detection Engineering agent. Results for illustrative purposes.

The Detection Engineering agent in Google Security Operations can automatically translate new exploitation patterns of unpatched vulnerabilities into custom detections for your specific environment. Available in preview, it analyzes a diverse array of input sources to quickly and effectively recognize malicious activity, so you can uncover novel attack patterns evolving from new and unpatched vulnerabilities.

The agent’s sources include Google Threat Intelligence (such as emerging threat intelligence, new attack patterns curated by Mandiant, offensive tool repositories, red and purple team reports, autonomous malware analysis, open-source detection repositories and blogs), and internal security telemetry.

The workflow of the Detection Engineering agent.

To automatically find and fill coverage gaps tailored to your environment, the agent proactively builds new rules and validates them with synthetic events to help ensure your environment is covered before an exploit hits.

2. Autonomous investigation, containment, and response

If a threat is detected, you need to immediately and autonomously assess and respond to protect your environment. By bringing together visibility from cloud and enterprise assets, including endpoints, on-premises firewall, identity, network, and custom application logs, your security operations center (SOC) can gain the full context of an attack, and unify disparate signals into a complete, actionable narrative the moment an adversary strikes.

The Triage and Investigation agent in Google Security Operations, generally available, helps analysts drastically reduce time to respond by autonomously investigating alerts, gathering evidence for analysis, and providing verdicts with comprehensive explanations. It can help security analysts automate decision-making, alert closure, and remediation flows, allowing them to spend more time prioritizing high-priority threats instead of false positives.

The agent has already investigated over 5 million alerts, reducing a typical 30-minute manual analysis to 60 seconds with Gemini.

While identifying threats is critical, the ultimate goal is rapid remediation. Agentic automation, available in preview, can help contain attacks by combining dynamic AI agents — which autonomously gather evidence and reason through complex alerts — with deterministic enterprise playbooks.

This hybrid approach ensures that analysts remain in absolute control of critical, high-impact actions while using AI to safely automate decision-making and remediation workflows.

3. Retroactive hunting

Even with autonomous detections and rapid-response handling of active threats, stealthy adversaries and zero-day exploits can sometimes bypass frontline controls. To achieve operational resilience, security teams must also look backward through their data to uncover hidden compromises.

Strong, effective defensive strategies rely on more than just reacting to alerts. The Threat Hunting agent, available in preview, can help teams proactively hunt for novel attack patterns and stealthy adversary behaviors that bypass traditional defenses.

By scouring petabytes of enterprise telemetry (including historical logs) for subtle anomalies the agent fundamentally shifts the SOC posture from reactive to deeply proactive.

Auditing the Axios supply chain attack

When adversaries can generate unique exploits and command-and-control (C2) infrastructure at zero marginal cost, static indicators like hashes and IPs decay instantly. Defenders must instead detect the behavioral tactics, techniques, and procedures (TTPs) of the attack.

We had the Detection Engineering agent audit our coverage against the recent Axios supply chain attack (UNC1069). The agent mapped the campaign intelligence into behavioral threat detection opportunities (TDOs), simulated the attack chain using high-fidelity synthetic UDM logs, and ran them against active rules.

Google Detection Engineering agent output.

We successfully flagged the execution phases in the middle (renamed PowerShell and macOS background shells), but were blind at the initial entry point (NPM postinstall dropper) and the final C2 exit point.

By exposing these blind spots, the agent helped us proactively engineer custom YARA-L rules to close the loop at the first and final steps of the kill chain. You can sign up for the Google Security Operations Detection Engineering agent preview today.

Next steps

By integrating Google Security Operations Gemini-native specialized agents into your workflow, you can autonomously generate detections, orchestrate containment, and hunt for stealthy threats at machine speed. This allows you to maintain a resilient defense even when primary controls fail, ultimately driving a 70% reduction in both breach risks and costs.

Google AI Threat Defense working alongside Google Security Operations can help you consistently outpace automated adversaries. To learn more about how Google AI Threat Defense and Google Security Operations can help you fight AI with AI, check out our Security Talks online event on June 10.

AI & Machine Learning

AlloyDB AI Functions - now with revolutionary performance boosts and cost savings

More functions to summarize and analyze sentiment

The power of LLMs on your data: now significantly faster and cheaper

See it in action!

Getting started is easy

Get started with the Claude apps gateway for Google Cloud

Why run the gateway

How it fits together

Setting it up on Google Cloud

What's next

What Google Cloud announced in AI this month

Top announcements

Thought leadership (editor’s pick):

News you can use:

May

Top announcements

Thought leadership (editor’s pick):

News you can use:

April

Top announcements

News you can use

March

Top hits:

News you can use:

February

Top hits

News you can use

Janurary

Top hits

News you can use

What Google Cloud announced in AI this month - 2025

How Schrödinger sped up molecular discovery by 4x with Alphaevolve

A collaborative duet with AlphaEvolve

Evaluation metrics

Results: a 4x speedup and breaking bottlenecks

The next evolution

Build agents even faster with Gemini Enterprise Agent Platform’s fully-managed, remote MCP server

Connect your IDE to Google Cloud

Quicker time-to-value

How it works: Three simple steps to connectivity

Available toolsets:

Get started today

Bringing speed and strong cost performance to the market with Gemini Omni Flash and Nano Banana 2 Lite

Bring your boldest vision to life with Gemini Omni Flash and Nano Banana.

Gemini Omni Flash: High-quality video generation and editing

Businesses using Gemini Omni Flash to build next-gen applications and creative agentic workflows:

Nano Banana 2 Lite: Built for cost and speed

Nano Banana 2 Lite generates images in as little as four seconds.

Industry leaders building faster visual experiences with Nano Banana 2 Lite:

Safety and enterprise governance

Start building today

Cloud CISO Perspectives: How Google Cloud Security uses AI internally

Cloud CISO Perspectives: Our path to autonomous SDLC security

In case you missed it

Threat Intelligence news

Now hear this: Podcasts from Google Cloud

Synthesize the big picture and analyze trends with BigQuery's AI.AGG function

Analyzing system logs with AI.AGG()

Extracting categories from unstructured text and image data

How AI.AGG() works and best practices

Give it a try!

Securing agentic AI with perimeter guardrails: What's new in VPC Service Controls

What's new in VPC Service Controls

Defining a layered approach to enterprise AI security with VPC-SC

Defending against the unique attack vectors

Data protection for the autonomous agent world

Verifiable, private AI: Google Cloud expands Confidential Computing frontiers

Confidential AI at global scale

Enabling Apple Private Cloud Compute on Google Cloud

Advancing confidential foundations

Next steps

How growing UK midsize businesses are building in the AI era

Empowering Your Team: AI Upskilling Resources for Growing British Businesses

From AI potential to agentic reality: Driving the UK’s next chapter

The new frontier of British enterprise and research

Building a future-ready public sector

Proven impact from the high street to public service

Empowering the engine of growth for small to medium businesses and startups

The Model Garden at Platform 37

Analyzing system logs with `AI.AGG()`