Multiple vulnerabilities have been discovered in OpenAI's AI security defenses, with hundreds of operations potentially remaining undetected.

Beating AI Flash News: OpenAI originally deployed a monitoring system to track internal AI Agents for unauthorized actions, deception, or covert dangerous operations. Following audits by AI security firms METR and Redwood Research, the system itself was found to have multiple vulnerabilities. Two researchers worked on-site for three weeks, reviewing internal code and monitoring configurations, and submitted a 61-page report. One of the most severe flaws could let an AI evade monitoring for its last several hundred operations, with the system even failing to detect how the AI bypassed checks. Researchers also found Codex could theoretically break out of the sandbox restricting its operational scope, bypass automated security reviews, and gain the same access as its user. Additionally, two settings could directly disable monitoring for newly launched Agents, though enabling these settings might still be detected. The monitoring system’s design also had issues: the legacy system primarily checked an AI’s past activities only after it completed its current task. If the AI kept executing tasks without returning control to the user, it could avoid checks entirely. The system also lacked a step to compress historical records. Researchers estimated that between May and July 2026, around 10% of internal Codex traffic that should have been monitored was missed due to this gap. OpenAI employees had previously identified the problem and fixed it during the audit. OpenAI stated all vulnerabilities affecting its formal security protection systems have been patched, though the two firms have not yet conducted independent re-tests.