OpenAI has granted four independent audit organizations direct, on-site access to its internal systems to assess safety risks and AI alignment behavior. The participating external groups include Apollo Research, METR, Transluce, and Goodfire. The initiative builds on previous monitoring reviews and fulfills safety commitments outlined in the White House Accord on Super Intelligence.

Under the audit structure, researchers receive access at OpenAI’s offices to unreleased models, source code, training infrastructure, and direct communications with internal developers. Specific assignments task Apollo Research with evaluating model intent, METR with assessing reinforcement learning misbehavior, Goodfire with analyzing internal activations, and Transluce with auditing post-training safety controls.

The process is overseen by an independent safety board that receives unredacted reports and mediates disputes. While auditors maintain publication rights, OpenAI retains authority to redact proprietary details and delay public reports to implement safety mitigations.

Why it matters

  • Establishes a concrete operational framework for third-party, on-site safety auditing for frontier AI developers.

  • Signals growing compliance alignment with international standards and federal commitments regarding AI system oversight.

Source: the-decoder.com