Decision integrity is the security property that a system’s outputs are not improperly manipulated. In practice, it means a ranking, score, recommendation, or approval should reflect the system’s intended logic, not hidden influence from an attacker or corrupted input. This matters any time software helps make business decisions, especially when the output affects access, hiring, payments, or risk scoring.
In cyber security, decision integrity is often threatened by prompt injection, data poisoning, tampered rules, or user input that is treated like instructions. For example, an AI screening tool that reads resumes or chat transcripts may be pushed toward a bad recommendation if the content contains directives the model follows instead of ignoring. Defenses include treating outside content as untrusted, separating instructions from data, validating outputs, logging decisions, limiting model authority, and requiring human review for sensitive actions. The goal is simple: external input should not be able to quietly steer the final decision.