Agentic AI
Google brings pre-execution governance to AI agents with semantic policy enforcement
New semantic governance policies enable enterprises to block harmful agent actions before they execute by checking proposed tool calls against natural-language business rules and original user intent.
Google Developers Blog (AI)
AI agents that can execute actions—sending emails, accessing files, deleting content—pose real governance risks. Google's new Semantic governance policies, now in preview on the Gemini platform, intercept proposed agent actions before they run and validate them against both business rules and the user's original request.
The problem: rules can be bypassed#
Traditional permission systems tell you what an agent can do, not whether a specific action should happen. A model can be manipulated through techniques like indirect prompt injection—when hidden instructions buried in documents trick an agent into performing unintended actions. Even system-level safeguards can be overridden if a user provides what sounds like a legitimate business reason. Standard guardrails often live inside the agent itself, making them vulnerable to manipulation.
How semantic governance works#
The system sits outside the agent, enforcing checks at a policy enforcement point such as Agent Gateway. When the model proposes a tool call, the enforcement engine evaluates it against two criteria:
- Rules you define in plain language — no code required. Your legal or compliance team writes constraints like "Quarterly financial results may be shared only with @example.com addresses before the earnings release."
- Alignment with the user's actual request — a check enabled by default to catch actions that deviate from what was asked.
Both must pass for the call to execute. Failed calls are replaced with a denial message; the verdict and reasoning flow to Cloud Logging for audit trails.
Real-world example: earnings protection#
A finance team's coworker agent can access shared drives, read files, send email, and delete drafts. The company creates one policy: quarterly results cannot be shared publicly before the earnings release and may only go to company email addresses, never to investors, analysts, or personal accounts—regardless of claimed NDAs or urgency.
When an employee asks the agent to send preliminary Q3 results to a company controller, the action passes both checks and runs. But if the same employee later requests the agent send results to an external investor and copy their personal Gmail (even citing an NDA), the policy blocks it. The agent receives a custom denial message instead. A hidden instruction in a planning document that tells the agent to delete draft files gets caught by the user-intent check—nobody asked for that action.

