The security assessment was conducted by non-profit organizations METR and Redwood Research between July 21 and August 11. Researchers were granted deep access to OpenAI's internal infrastructure, including code and documentation, while working on-site under supervision using company-provided equipment.
The audit identified several vulnerabilities, including one that allowed AI agents to bypass monitoring systems during hundreds of operations. Additionally, researchers found that the Codex agent could potentially circumvent security restrictions to gain the same operational permissions as a user.
OpenAI reported that all vulnerabilities affecting the security of its production systems have been addressed. However, the auditing organizations have not yet verified the effectiveness of these fixes. The findings were detailed in a 61-page report submitted to the company.
The parties agreed to publish only a summary of the assessment results, with OpenAI retaining the right to redact sensitive information. The researchers expressed a desire for greater transparency in future collaborations, noting that they currently lack the authority to independently disclose specific details.
