Google Gemini: security advisories, deprecations and releases

LLMJacking: threat actors monetize hijacked cloud LLM resources

Mandiant details LLMJacking campaigns where attackers steal cloud IAM tokens and LLM API keys, then run expensive inference workloads that can rack up six-figure bills in hours.

CVE-2026-40217

Agent identity, sandboxing and policy enforcement via Agent Gateway

Every Gemini agent receives its own attested identity, fine-grained RBAC, full audit trails and runs inside a sandbox with traffic controlled by Agent Gateway policy.

Agent Development Kit (ADK) deprecated on DigitalOcean

The Agent Development Kit, gradient-adk Python package, and gradient CLI are deprecated. No new agents can be deployed, but existing ones continue to run and be billed.

Reinforcement Learning best practices for Gemini customization

End-to-end walkthrough covering when to apply RL fine-tuning and how to avoid reward hacking.

Developer’s guide to MCP apps in Gemini

Step-by-step instructions for building multi-step, visual agents that call external tools without repeated chat turns.

Three ways to audit LLM evaluators

Framework to detect bias and drift in automated judges that score agent output quality.

Secure Gemini Enterprise to Snowflake integration with governed Cortex Agents

Shows how to connect Gemini Enterprise to Snowflake data using OAuth user impersonation, Cortex Agents, and MCP servers so AI agents generate and run SQL inside Snowflake's security boundary.

Graph workflows in ADK: fan-out, routers, human review, parallel workers

Explains how to build refund workflows with ADK using fan-out/fan-in, deterministic and agent routers, human-in-the-loop pauses, parallel workers, and dynamic nodes.