AI guardrails expire before actions run

AI guardrails can approve an action correctly and still be wrong when execution arrives: a five-system study found 3.4% to 24.7% of valid approvals expired after an eight-step delay. The study of verdict staleness cut expiry to 0%–1.8% with a freshness-bounded shield that estimates an approval’s safe lifetime from its margin and recent state volatility. For teams building agent control planes, a pass/fail guard is incomplete unless it carries a time-to-live and rechecks state before side effects.