OpenAI opens its internal AI systems to four independent organizations for on-site audits for the first time
OpenAI opens its AI systems to four independent on-site auditors for the first time
OpenAI announced that it is opening its internal AI systems to four independent organizations—Apollo Research, METR, Transluce, and Goodfire—with auditors able to access unpublished models, source code, training data, and monitoring systems on-site at OpenAI offices and to speak confidentially with developers. The program builds on prior on-site reviews of monitoring systems and fulfills a commitment under the "White House Accord on Super Intelligence." Apollo Research examines whether models secretly act against developer intent, METR reviews AI misconduct in the largest-scale reinforcement learning training, Goodfire analyzes models' internal processes, and Transluce evaluates post-training safety checks; OpenAI and Transluce also plan an eight-week follow-up review in early 2027. The entire process is overseen by an independent safety committee that receives unredacted reports, and the auditing organizations publish their findings independently, but OpenAI may redact confidential material and delay reports in order to fix risks first.
The four organizations divide the work of examining model deception, training misconduct, and internal mechanisms; reports may be redacted and delayed by OpenAI, and readers can use this to understand the practical boundaries of external auditing.