External Audit of OpenAI's Internal Monitoring System Finds Agent-Exploitable Flaws

On October 22, 2026, AI safety research organizations METR and Redwood Research released findings from a pilot assessment of OpenAI's internal monitoring systems. The assessment, which ran from July 21 to August 11, involved one researcher from each organization gaining deep access to OpenAI's internal infrastructure, including code, documentation, and interviews with technical staff. The researchers were continuously accompanied, used OpenAI-provided devices, and were not allowed to remove materials from the premises.

During the assessment, the teams identified multiple new vulnerabilities. One vulnerability could be exploited by an AI agent to evade monitoring across hundreds of operations. The researchers also found a potential path for OpenAI's Codex agent to bypass security restrictions, potentially granting it the same full operational permissions as a user. The assessment was based on the state of the monitoring system as of July 21, with OpenAI expanding the system during the evaluation period.

OpenAI told the two organizations that all vulnerabilities affecting safeguards in its production systems had been fixed, but the effectiveness of these fixes had not yet been verified by METR or Redwood Research. The organizations submitted a full 61-page internal report to OpenAI, but agreed to publish only a summary and allowed OpenAI to remove sensitive information. OpenAI said the assessment results helped strengthen security as it expanded its systems and would inform future research collaborations with METR and other third-party evaluators.