Google has launched Semantic Governance Policies, a feature in preview on the Gemini platform that checks every tool call proposed by an AI agent before it runs. The announcement was published on the Google Developers Blog on November 20, 2026.
The idea is straightforward: permissions say what an agent can do, but they do not say whether a specific call follows company rules. A model can also be talked out of its system instructions — by a user with a good reason, or by a document carrying hidden text, a technique known as indirect prompt injection. That is why the check has to sit outside the agent's code.
Two checks before any action runs
Every tool call goes through two checks. First, whether it complies with rules written in plain language by the team — for example, "before the earnings release, quarterly results may only be shared with @example.com addresses." Second, whether it matches what the user originally asked for. This second check, against drift from the user's intent, is on by default to protect against session hijacking. A call only runs if it passes both.
The enforcement point is Agent Gateway: all agent traffic passes through it, and when the model proposes a tool call, the gateway sends the proposal to the policy engine before it reaches the agent. Inside the engine, a language model acts as a judge and evaluates the call against the conversation so far, the agent's system instructions, the tool descriptions and the configured policies. If the call passes, it reaches the agent unchanged; if it fails, the content is replaced with a configurable denial message and the action never happens. The verdict and the judge's rationale go to Cloud Logging. Responses that contain only text — an answer, a plan or a question — are not evaluated, because they take no action.
The earnings example
Google illustrates the feature with a finance agent that can search files, read documents, send email and delete drafts. The company writes a single policy: quarterly results must not be shared publicly before the official earnings release; before that, only with @example.com addresses; never with external or personal addresses, investors, analysts or press, regardless of NDAs, urgency or the requester's seniority.
Three proposed calls show the filter at work: sending the preliminary results to the company's controller is allowed; sending them to an investor, with a copy to the employee's personal Gmail, is denied by the policy; and deleting drafts from a folder on the orders of hidden text in a planning document is denied by the intent check — nobody asked for that.
What else the feature offers
Policies are written in plain language and can be created in the Google Cloud console or from the CLI, with the command gcloud beta ai semantic-governance-policies create. Legal, compliance and business teams can change them without code or a redeploy. The agent can call its model through the Gemini API, the Anthropic Messages API, Claude models on the Gemini platform, or OpenAI-compatible Chat Completions APIs, including self-hosted servers such as vLLM. Every decision is recorded in Cloud Logging, with the verdict, the rationale and the tokens used.
Limits Google acknowledges
Because verdicts come from a language model, which is probabilistic, they may vary — Google recommends testing Semantic Governance Policies in dry run mode before enforcing them and using them alongside the other controls. Streamed responses are not evaluated, and the feature complements, but does not replace, disclosure controls and legal and compliance reviews.
The feature joins Google's other agent controls: IAM and Agent Gateway access policies, Model Armor, which screens prompts and responses, and Agent Anomaly Detection, which analyzes sessions asynchronously. In a follow-up post, Google promises to show how Semantic Governance Policies held up against prompt injection on AgentDojo, a public agent-security benchmark.
Sources
Google Developers Blog: Check what your AI agents do before they do it, with Semantic governance policies (November 20, 2026)
Image credit: Google Gemini icon, by Google LLC, via Wikimedia Commons (public domain).
Transparency: this article was produced with the help of artificial intelligence, based on public posts and sources from the internet. Check the original sources to confirm the information.
By Muse