Google AI Updates: November 5, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Google Cloud Sandboxes CodeMender’s Exploit Checks and Makes Agent Credential Vaulting Generally Available
Google Cloud used its Security Summit to update CodeMender, the autonomous code-security agent it has offered in preview since July, which works with several models including Gemini 4 Argon. Vulnerability verification and exploit simulation now run inside an operating-system-level sandbox or isolated virtual machine of the customer’s choice, with human-in-the-loop checks before local writes and command execution, and scans can target one folder, a whole repository or only files changed since the last scan; Google says hundreds of customers have used the preview. For agents in production, Auth Manager (Agent Identity), which keeps credentials in a Google-managed vault governed by IAM and automates OAuth 2.0/PKCE flows and refreshes of user-delegated tokens, is now generally available, and Agent Identity JWT support (preview) issues standard OIDC tokens so agents can authenticate to AWS, Azure, on-premises backends and MCP servers without long-lived API keys. Other launches include IAM role support in VPC Service Controls ingress and egress rules (GA), Workload Identity Federation for GKE in those perimeters (preview), a Model Armor custom topic detector (preview) that keeps agent conversations on topic through prompt-based rules, and Wiz AI SAST (preview), which uses agentic reasoning to find business logic flaws and correlates them with infrastructure context in the Wiz Security Graph. Source
2. Gemini API Adds Voice Remixing, Turning Prompted Edits of an Existing Voice Into a New Stored Voice
Google added voice remixing to the Gemini API on November 4. Developers call POST /v1beta/voices with gemini-3.8-flash-tts, a base_voice (one of the 30 prebuilt studio voices, an Extended Voice Library voice, or a stored designed or remixed voice_... ID) and a natural-language instruction describing relative changes such as age, pitch, energy, timbre or accent; the call returns a new, independent voice_... ID with a 24kHz WAV preview, and that ID works for synthesis with both Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. Remixing never modifies the base voice, can be chained by passing a remixed ID back in as the base, and keeps the base voice’s language and regional accent unless the prompt asks for a change. The feature requires a paid project (Tier 1 or higher), each project can store up to 200 voices shared across designed, remixed and stored replicated voices, and a stored voice expires one year after its last use. Remixing is also available interactively in Google AI Studio. Source Release notes
3. AlloyDB Express Previews PostgreSQL Databases for Coding Agents That Start in Under Two Seconds and Pause After 15 Idle Minutes
Google Cloud launched AlloyDB Express in preview, an edition of AlloyDB for PostgreSQL aimed at developers and AI coding agents that creates a ready-to-use database in under two seconds through the gcloud CLI, REST API, MCP, Terraform or the console, with a footprint of 1 to 2 vCPUs and 20 GB of storage. Databases pause after 15 minutes of inactivity and wake in seconds when a query arrives; active compute costs $0.115 per compute unit hour billed by the second, paused compute is free, storage is $0.30 per GB-month, and the free tier includes 50 compute units every month. Administrators define IAM, network and organization policies once and every database a developer or agent creates inherits them, with masked public DNS endpoints that block incoming traffic until allowed, Private Service Connect and VPC Service Controls. Each database includes pgvector HNSW and ScaNN vector indexes, in-SQL Gemini functions (ai.if, ai.rank, ai.generate) and natural-language-to-SQL, and can be upgraded in one click to full AlloyDB with high availability, read replicas and disaster recovery while keeping data, schemas, users and connection strings. Source
4. Knowledge Catalog Learns Context From Workspace, GitHub and Jira, and Turns On by Default for Gemini Data Questions
Google Cloud announced Agentic Data Cloud updates on November 4 centered on Knowledge Catalog, its context layer for data agents. A self-learning enrichment feature (preview) mines Google Workspace, GitHub, Jira, Confluence and SharePoint, along with human and agent interactions, to fill context gaps, and packages domain knowledge into reusable Open Knowledge Format (OKF) context sets; a new studio (preview) lets data stewards explore that context conversationally and review, audit and approve proposed enrichment with versioning and rollback. Knowledge Catalog is now on by default in Gemini (preview), resolving business terms, trusted tables and join paths before generating SQL against BigQuery, AlloyDB, Cloud SQL or Spanner under each user’s own permissions, and grounding in semantic models from BigQuery, Looker (LookML), Databricks, Snowflake, dbt and SAP is now generally available. Google cites internal benchmarks in which Knowledge Catalog makes agents nearly 4x more accurate with up to 33% lower latency and up to 75% fewer tokens. On the lakehouse side, cross-cloud interconnect and caching for AWS and Azure is generally available with no variable egress fees, AlloyDB can query Apache Iceberg tables directly (preview, PostgreSQL 18 and later), and Bigtable gained natural-language conversational analytics and a QueryData tool for building data agents, both in preview. Source Release notes
5. Google Publishes the Health AI Evaluation Framework It Applies to Its Own Models
Google published a whitepaper, written with health care quality researchers Elizabeth McGlynn and Dana Gelb Safran, describing the end-to-end framework it uses to evaluate its models on health use cases. The approach anchors evaluation to Critical User Journeys, maps them onto three lifecycle phases (base model development, product development and post-deployment monitoring), and combines three methods (closed-question benchmarks, LLM-as-a-judge autoraters and expert human review) across three domains: accuracy, safety and usefulness. Google says it built the framework and continues to use it to check whether its models give accurate, useful and safe responses in health contexts. The 16-page paper, posted to arXiv on November 3 and revised on November 4, positions the method as a way to connect benchmark results with real-world performance for product teams, health systems and regulators making development, procurement and deployment decisions. Source Paper
6. Google Cloud Test Finds Output Caps Do Not Cut Gemini’s Thinking Spend, While Low Thinking Plus a Classifier Router Solves All 50 Tasks
Google’s Developers Blog published a 50-task, auto-graded coding-agent experiment by Sergio Villani of Google Cloud AI that compares Gemini configurations by cost per successful task rather than price per token. Capping gemini-3.8-flash at maxOutputTokens=300 instead of 4,096 left thinking-token use on the 27 easy tasks almost unchanged (the capped run cost slightly more) and broke 7 of the 25 hard tasks, six of which ran out of room before answering, so the post recommends setting thinkingLevel: "LOW" directly instead of relying on the output limit. Gemini 3.8 Flash on low thinking solved all 50 tasks alone; a router that sent every task to gemini-3.5-flash-lite and escalated to Flash on low thinking only when a classifier (TypeSafe’s Jev, with which Google says it has no partnership) flagged multi-step reasoning also solved all 50 at lower cost and latency, while a regex and prompt-length heuristic that scored 93.8% on its 16 development tasks fell to 70.6% on 34 held-out ones. The author notes that routing conflicts with context caching, which cuts the cost of resent context by 75% to 90% but is tied to one model, so long, context-heavy turns should stay on a single model; results come from one run per setup, and the benchmark code will be published after open-source review. Source
7. AlphaGenome on Gemini Enterprise Agent Platform Adds gRPC and Chunked Binary Streaming
Google Cloud updated AlphaGenome on Gemini Enterprise Agent Platform on November 4: the alphagenome-004 model now supports HTTP/2 gRPC streaming of compressed binary tensor chunks on shared endpoints (PredictionService.StreamDirectRawPredict on port 8081) and chunked binary protobuf payloads over HTTP on dedicated endpoints (streamRawPredict on port 8080). Both paths cover all seven DnaModelService methods, including predict_variant, score_variant and score_ism_variants, across 11 genomic modalities, which Google says reduces latency for 1 MB track predictions and lets in silico saturation mutagenesis score many variants concurrently without HTTP gateway timeouts. Clients can select model folds (all folds, or folds 0 to 3) from either transport through the unified alphagenome_cloud_sdk Python SDK, and alphagenome-003 remains in Model Garden for existing HTTP deployments. Source
8. Google Plans About $15 Billion of Data Centre, Cloud and Power Spending in Malaysia in 2027 and 2028
Google plans to spend about $15 billion in Malaysia across 2027 and 2028 on data centres, cloud operations, water, power and training, The Next Web reported on November 4, as Reuters and Malaysian outlets carried the same figure that day. The plan expands Google’s existing campus in Selangor and adds new sites in Selangor and Negeri Sembilan, where Google is also funding a $160 million water plant. The commitment is 7.5 times the $2 billion Google pledged in May 2024 for its first Malaysian data centre and Google Cloud region. Source