A human review gate is a checkpoint that stops an AI agent from taking a sensitive action until a person approves, edits, or inspects it. The action might be sending an email, changing a record, sharing a file, or triggering another system. In practice, the gate turns an autonomous step into a supervised one, so the model can draft, recommend, or prepare work without being fully trusted to execute it.
In cyber security, human review gates reduce the blast radius of mistakes, prompt injection, and abused permissions. If an agent is tricked by malicious content or misunderstands instructions, the gate gives operators a chance to catch unsafe output before it causes damage. Defenders use these checkpoints for high-risk actions, paired with least privilege, logging, and clear approval rules. Attackers, in turn, may try to bypass or fatigue reviewers, so the gate works best when it is narrow, well-defined, and reserved for actions that truly need human judgment.