Deactivation criteria are predefined conditions that tell defenders when to disable, restrict, or roll back a system, feature, or permission set. In cyber security, they act as a safety switch: if risk signals cross a set threshold, the system is no longer trusted to operate at full capability.
For agentic AI and other automated services, deactivation criteria may include repeated prompt injection attempts, unsafe tool calls, abnormal output patterns, policy violations, or signs of abuse. They matter because an AI system with access to data, workflows, or external tools can turn a mistake into real impact. Well-designed criteria support rapid containment, human review, logging, and least-privilege response. Attackers may try to stay below these thresholds, while defenders use them to stop escalation before damage spreads.