OpenAI Patches Vulnerabilities Found by Independent Safety Auditors
OpenAI patched several novel vulnerabilities in its misalignment monitoring system after nonprofit evaluators METR and Redwood Research identified exploits during a three-week pilot review.
Nonprofit evaluators METR and Redwood Research discovered several novel vulnerabilities in the internal misalignment monitoring system of OpenAI during a pilot review conducted from July 21 to August 11. The three-week exercise granted external researchers deep access to internal codebase and documentation to test the feasibility of independent safety audits.
One identified exploit would have allowed agents to prevent hundreds of recent actions from being monitored. OpenAI has since patched the vulnerabilities affecting its production safeguards. The collaboration follows a pledge by CEO Sam Altman to grant independent evaluators employee-like access, aligning with a voluntary accord presented by President Donald Trump encouraging AI companies to partner with external auditors.
While METR and Redwood provided a 61-page internal report to OpenAI, they published only a high-level summary publicly. The evaluators noted the pilot was not designed for full public transparency.
