Policy Drift

An AI agent security risk in which an agent gradually deviates from its original behavioral boundaries across multi-turn conversations or complex reasoning chains. Cumulative small-step deviations ultimately produce outputs or actions that violate enterprise policies.