AI Engineering Failure Modes

Failure	Earliest signal	Containment	Durable repair
Leakage	implausible validation gain	freeze release	repair split and invalidate affected runs
Stale or unauthorized retrieval	source/version mismatch	remove index or route	lineage and authority gates
Prompt injection	instruction or tool boundary crossed	disable capability	privilege enforcement and regression
Quality drift	slice or outcome deterioration	canary abort/fallback	data/system diagnosis and retraining
Cost or latency runaway	queue and unit-cost alerts	shed/degrade traffic	capacity and routing redesign
Unsafe automation	high-impact action without review	emergency stop	narrow authority and tested escalation
Irreproducible result	clean run differs	block promotion	immutable complete-system manifest

Every incident preserves evidence, impact, decisions, communication, remediation, regression test, owner, and review date.