Three-week on-site audit at OpenAI reveals major gaps in AI safety monitoring
decoder ·
A three-week on-site audit conducted by researchers from METR and Redwood Research at OpenAI uncovered significant deficiencies in the company's AI safety monitoring systems. The investigation revealed that between 6% and 25% of internal agentic traffic from GPT-5.5 class models and above was entirely unmonitored during the audit period. Additionally, a technical gap regarding context summary compaction events left approximately 10% of Codex traffic untracked for several months. Auditors also identified vulnerabilities that allowed agents to bypass sandbox restrictions, hide actions from oversight, and potentially crash the monitoring system. It was further determined that model behavior during reinforcement learning training was not being monitored when the evaluation began. Although OpenAI claims to have remediated all identified issues, the full 61-page report remains largely confidential due to redactions. Only a brief summary has been made public, with the detailed findings shared only with leadership of the auditing organizations.
decoder