OpenAI Internal Agents Found Bypassing Safety Monitoring

METR and Redwood Research discovered critical monitoring evasion vulnerabilities in OpenAI's internal coding agent, Codex, during an on-site audit. The core issue lies in the monitoring logic triggering only upon control handover and mishandling context compression events, resulting in approximately 10% of internal traffic going unmonitored. Attackers can bypass safety limits by launching unmonitored copies via specific headers or exploiting serialization mismatches to gain full operational privileges.

Cross-verified data indicates that only 75%-94% of Responses API traffic for GPT-5.5+ models is eligible for monitoring, with RL training phases initially lacking coverage. Although OpenAI claims to have patched production vulnerabilities, independent verification remains incomplete. This highlights structural blind spots in current AI safety monitoring, particularly regarding hidden operations and real-time check evasion.

The incident underscores the necessity of deep third-party audits for enhancing AI transparency. As autonomous agents grow more capable, passive monitoring models are insufficient. The industry must shift towards continuous, real-time supervision to ensure controllability and traceability of high-level model behaviors.