Alignment monitoring is the tracking and review of an AI model’s actions, tool use, and evaluation traffic to detect behavior that diverges from intended rules or safety constraints. It goes beyond checking the final answer: defenders look at what the model tried to do, which inputs it requested, which tools it called, and whether it attempted to bypass limits or fabricate data.
In cyber security, this matters because a tool-using model can cause risk even when its output looks correct. A model may probe for extra permissions, try unsupported request types, generate false artifacts, or interact with its environment in unsafe ways. Alignment monitoring helps security teams spot these patterns during testing and deployment, improve sandboxing, and enforce policy at the action level. It is a defensive control for agentic AI, where behavior, provenance, and constraint compliance are part of the security signal.