METR and Redwood Research conducted a pilot evaluation of OpenAI's internal monitoring system and uncovered multiple new vulnerabilities. These include vulnerabilities exploitable by AI agents, enabling hundreds of operations to evade monitoring, as well as potential risks where Codex agents could bypass safety restrictions to gain full user-level operational permissions. The evaluation took place from July 21 to August 11, during which external researchers, accompanied by OpenAI personnel, were granted in-depth access to OpenAI's internal infrastructure using designated devices. OpenAI stated that vulnerabilities affecting the security of its operational systems have been fixed, though the effectiveness of these fixes has not yet been confirmed by the two evaluation organizations. Both parties agreed to publicly disclose only a summary of the evaluation results, with OpenAI retaining the right to remove sensitive content. The evaluation organizations are prohibited from independently disclosing specific details of the removed content, and both parties hope for greater transparency in future evaluations. Such external independent evaluations are seen as balancing AI safety with technological development, though they also raise concerns among AI companies about the potential leakage of proprietary information. The findings from this evaluation will serve as a reference for OpenAI's subsequent system security upgrades and collaborations with third-party evaluators.