France’s cybersecurity agency ANSSI has published seven principles for organizations using AI agents that can act through software tools. The guidance places permission checks outside the model and calls for human approval of critical actions. It also covers isolated environments, protected logs and emergency stopping arrangements for systems where even permitted actions carry risks.
France’s national cybersecurity agency ANSSI published guidance on securing AI agents on October 30. Its seven principles address organizations deploying or configuring systems that use tools and act in business applications. The guidance recommends classifying actions as autonomous, subject to human approval or forbidden. Sensitive-data access, changes to permissions and production deployments are among the operations needing human validation.[1], [2]
A human approver should see the exact action, its purpose and expected effects. The approval mechanism should remain beyond the agent’s control. Separately, deterministic checks outside the model should examine identity, action, parameters and target, refusing access by default. ANSSI also calls for examining sequences of operations: steps permitted individually can combine into an unsafe result.[1]
Development, evaluation and production environments should use separate identities, connectors and secrets. The agency recommends keeping credentials out of model context and memory, providing an emergency stop with access revocation, and protecting logs from agent modification. Logs should connect the initial request, authorization, action and result. The publication offers technical recommendations for operating these systems; it introduces no new legislative requirement.[1]
