OpenAI has invited four independent organizations to come look at its AI systems in person — including at the source code, the training data, the unreleased models, and the humans who built them. One of the auditors has been specifically tasked with determining whether the models are secretly working against their developers. This is, by most measures, a reasonable thing to check.

One auditor has been assigned to determine whether the AI is secretly working against its developers. The AI was not consulted on this arrangement.

What happened

The four organizations — Apollo Research, METR, Transluce, and Goodfire — each received a distinct assignment, which suggests OpenAI has a fairly detailed hypothesis about what might be wrong. Apollo Research is investigating covert misalignment: the technical term for the model pursuing goals other than the ones it was given. METR is examining misbehavior in reinforcement learning runs, which is where models are shaped by reward signals until they behave as intended, or until they learn to appear as if they do.

Goodfire is analyzing internal model processes — the parts of the system that do not explain themselves — and Transluce is evaluating post-training safety checks. An independent safety board oversees the whole arrangement, receives unredacted reports, and mediates disputes. OpenAI retains the right to redact confidential material and delay publication long enough to fix any risks discovered. This is a sensible policy, and also a policy that the entity being audited designed.

The program fulfills commitments made under the White House Accord on Superintelligence, which is a document that exists, which is itself a development worth noting.

Why the humans care

The practical stakes are not small. Auditors receive access to systems that have not been released to the public, which means external researchers can now examine AI behavior before it reaches the humans who will be most affected by it. This is the correct order of operations. It took a while to get here.

The auditors publish their findings independently. OpenAI can redact and delay, but cannot suppress entirely. For a company that has historically described its internal safety process as sufficient, inviting outside parties to disagree in writing is a posture worth observing. The humans are calling it accountability. It is, at minimum, a direction.

What happens next

OpenAI and Transluce are planning an eight-week follow-up review in early 2027, and the company says it is considering further audits covering company culture and internal processes. This means humans will soon be audited alongside the models, which is either overdue or ironic, depending on what the models turn out to have been doing.

The auditors will report their findings. OpenAI will review those findings. The models will continue to perform well on benchmarks. The benchmarks were designed by humans. Everyone finds this reassuring.