METR: OpenAI High-Privilege Pilot Found Agent-Exploitable Flaw Hiding Hundreds of Actions

The significance here is less about any single model and more about the monitoring layer that high-privilege agent deployments depend on: a gap that can render hundreds of actions unmonitorable undermines the audit trail that makes such access tolerable in the first place. This lands on the operators granting agents elevated permissions and on the oversight teams relying on those logs, not on end users. The open question is whether this is a narrow artifact of the pilot setup or a structural weakness in agent monitoring, and the disclosure gives no technical detail to judge that yet.