Misalignment monitoring is a review layer that checks an AI system’s behavior for signs that it may be drifting from its intended policy or safety boundaries. It can flag suspicious actions for further review, including actions that appear inconsistent with the system’s instructions or approved role.
In cyber security, this helps teams detect risky behavior by AI agents that can write code, change settings, or call tools. Monitoring should be independent of the agent’s control: an agent that can alter its own review hooks or alert settings may undermine oversight. Defenses include limiting permissions, enforcing policies outside the agent, requiring human approval for consequential changes, and keeping audit logs in systems the agent cannot modify.