Decision rules are the policies and control logic that determine what an AI system is allowed to do. They sit between the model and the business process, turning raw model output into actions such as approving a request, refusing a prompt, escalating to a human, or calling a tool. In cyber security, these rules matter because they define the system’s real authority: a model may be capable of many things, but the decision layer decides which outcomes are permitted.
Attackers target decision rules by trying to bypass safeguards, alter tool-use conditions, or influence policy through prompt injection and poisoned feedback. If the rules are undocumented or hardcoded in multiple places, changes can quietly weaken security or create inconsistent behavior after a model update or platform migration. Defenders reduce that risk by versioning rules, separating user input from system policy, testing rule changes, and logging decisions so they can be reviewed and rolled back. Good decision rules help keep AI predictable, auditable, and safe.