i
News
News · 2026-11-19

OpenAI lets four independent auditors examine its AI systems

Dataist Dataist

OpenAI has, for the first time, allowed four independent auditors to examine its AI systems on site

Cover: OpenAI lets four independent auditors examine its AI systems

OpenAI has, for the first time, given four independent organizations access to its internal systems so they can examine safety risks and model behavior on site. Depending on the assignment, Apollo Research, METR, Transluce and Goodfire will be able to work with unreleased models, source code, training data and monitoring systems, as well as speak confidentially with developers. The program builds on an earlier review of OpenAI’s monitoring systems and fulfills the company’s commitments under the White House agreement on superintelligence.

OpenAI lets four organizations examine its internal AI systems

What the auditors are examining

The organizations have different assignments. Apollo Research is investigating whether models try to act covertly against their developers’ intentions. METR is studying improper AI behavior during the largest reinforcement learning cycle, a method in which models learn from rewards. Goodfire is analyzing models’ internal processes, while Transluce is assessing safety evaluations after training.

OpenAI and Transluce also plan to conduct an additional eight-week review in early 2027.

Who oversees the process

An independent safety board oversees the audits. It receives unredacted reports and helps resolve disagreements. The organizations publish their findings independently, but OpenAI can withhold confidential information and delay publication to address identified risks first.

The company expects some friction at the outset. OpenAI is already considering new reviews, including ones focused on corporate culture and internal processes.

Dataist’s take: Access to internal materials lets auditors examine systems more deeply than a review based only on published results. But OpenAI retains the right to edit reports and delay their publication, so the review’s independence will also depend on how those limits are applied in practice.

Daily AI news

Every day we pick what actually matters in AI and explain it plainly — no hype, no filler. Subscribe if you want to follow where the industry is going.

Only what matters — every day

Follow on X