Brief IA
›

Four Organizations Audit OpenAI's Internal Systems

AI Tools·Tom Levy·

Four Organizations Audit OpenAI's Internal Systems

Four Organizations Audit OpenAI's Internal Systems
Key Takeaways
1Four independent organizations access OpenAI's internal on-site systems.
2An independent council receives full reports and mediates disputes.
3OpenAI and Transluce plan an eight-week follow-up starting in 2027.
Why it matters : Audits provide access, depending on the missions, to unprecedented models, source code, and training data.

OpenAI welcomes four independent organizations for the first time to directly audit its internal AI systems on-site. Their investigations focus on the safety and behaviors of the models.

OpenAI can defer auditor publications

An independent security council oversees the supervision, reviews the reports in their entirety, and acts as a mediator in case of disagreements. The organizations publish their findings autonomously; however, OpenAI retains the right to redact confidential elements or postpone a publication while addressing identified risks.

The company anticipates some friction at the outset. It also considers additional reviews focused, among other things, on its corporate culture and internal procedures.

The missions of the four auditors

OpenAI designates Goodfire, Transluce, METR, and Apollo Research as partners in the program. Goodfire investigates what happens inside the models, while Transluce examines the security controls implemented after their training. With the latter organization, OpenAI plans a new review lasting eight weeks, starting in early 2027.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

METR looks for inappropriate behaviors during the largest reinforcement learning training session, a technique based on rewards. Apollo Research, for its part, seeks to determine whether the models act against the developers' intentions without their knowledge.

Auditors can access unpublished models

In the company's offices, the accessible scope varies with the mission: unpublished models, source code, data used for training, or monitoring devices. Confidential interviews with developers are also possible.

This arrangement extends a previous review of monitoring tools already conducted on-site. It fulfills the commitments of the White House Accord on Super Intelligence.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.