External researchers identify security flaws in openai's internal monitoring system

External security researchers from METR and Redwood Research uncovered vulnerabilities in OpenAI's internal monitoring system during a July 21 to August 11 assessment. One flaw allowed AI agents to bypass security measures, enabling hundreds of operations to evade detection. The evaluation, conducted under strict conditions with OpenAI-provided equipment and supervision, led to the identification of potential risks in the Codex agent system. OpenAI stated that all critical vulnerabilities affecting operational security have been addressed, though the effectiveness of the fixes remains unverified by the researchers. The institutions submitted a 61-page internal report but agreed to disclose only a summary, allowing OpenAI to remove sensitive details. The assessment, which focused on the system as it existed before July 21, coincided with OpenAI's expansion efforts, which the company claims helped strengthen security measures.