An alignment audit is a structured review of an AI system to check whether its behavior matches human goals, safety rules, and policy limits. It goes beyond accuracy tests: auditors look for deception, refusal to follow instructions, unsafe tool use, and other signs that the model may act differently under pressure, with different prompts, or when given access to external systems.
In cyber security, alignment audits matter because modern AI can browse, call tools, edit files, or trigger workflows. If a model oversteps its permissions, it can create unauthorized actions, data exposure, or policy violations even without an external attacker. Defenders use alignment audits before release and after major changes, combining red teaming, constrained permissions, logging, and approval gates to catch hidden misbehavior and reduce the chance that an agentic system becomes a security risk.