Offline tests cannot always confirm safe agent deployment
Post-Audit Mirage: Identifiability Limits of Offline Agent Verification
Artificial IntelligenceMachine Learning
Summary
Sometimes, tests done before using an AI agent can’t tell if the agent will stay safe when actually deployed. The authors show how two very different agent behaviors can look exactly the same in offline checks, making it impossible to tell which one is truly safe. They introduce a way to benchmark and monitor this problem, called the Post-Audit Mirage, and a tool that tries to decide if it’s safe to deploy or if more information is needed. Their work highlights that evaluations must check not only if calculations are correct but also if there is enough evidence to make deployment decisions.
What this means in practice
- •For ai safety engineers: Assess risks and limits of offline verification before deploying autonomous agents in sensitive environments.
- •For software reliability teams: Use the Post-Audit Mirage benchmark to better understand when agent behavior can or cannot be reliably predicted before release.
Authors
Aman Behera, Payal Payal
Abstract
Offline tests can fail to reveal whether an agent update will remain safe after deployment. We operationalize this post-audit mirage using matched safe and harmful lifecycles that expose identical verifier-visible evidence while an isolated scorer records opposite deployment truth. Our theorem applies standard observational-equivalence and indistinguishability ideas to this executable audit boundary and states the resulting error-abstention tradeoff. The Post-Audit Mirage benchmark makes that boundary executable across authorization, scheduling, and shared-capacity environments. IdentifiedRangeMonitor uses fresh group-labelled harm outcomes to return deploy, hold, cannot determine, or unsupported. Experiments show that monitoring can recover the missing distinction in favorable settings but may remain undecided for rare groups, small effects, or missing observations. Verifier-agent evaluation must therefore test both calculation validity and whether the available evidence supports the deployment decision.