OpenAI has granted four external organizations direct access to its internal AI systems to conduct on-site safety and behavioral audits. The participating firms include Apollo Research, METR, Transluce, and Goodfire. This initiative fulfills commitments made by the company under the White House Accord on Super Intelligence.

Auditors receive access to unreleased models, source code, training data, and monitoring systems at OpenAI offices. They are also permitted to hold confidential discussions with developers. Apollo Research focuses on potential model misalignment, while METR examines misbehavior in reinforcement learning processes. Goodfire analyzes internal model operations, and Transluce evaluates post-training safety protocols.

The program is supervised by an independent safety board that mediates disputes and reviews unredacted reports. While auditors publish their findings independently, OpenAI retains the right to redact sensitive information and delay publication to address identified risks. A follow-up review with Transluce is already scheduled for early 2027.

OpenAI anticipates initial challenges with the audit process but is exploring further assessments. Future evaluations may extend to include reviews of the company's internal culture and operational procedures.