Topics: Evaluation and testing, Monitoring and control, Alignment research
OpenAI starts outside expert reviews of its AI alignment and safeguards
In short
OpenAI says it has signed agreements with four outside groups to review how its AI models behave and how well its safety measures work. The groups are Apollo Research, METR, Transluce and Goodfire. The post describes plans and rules. It does not report any model misbehaving.
What happened
OpenAI published the post on November 16, 2026. It calls the effort "embedded assessments." In these, outside experts get deep access to a company's internal systems and study them over a long period.
OpenAI says the experts may see model checkpoints (saved snapshots of a model during training), source code, training and evaluation data, monitoring systems and technical documents. They may also talk with the staff who did the work. The first assessments cover alignment, meaning whether a model does what its developers intend, and safeguards, during training and internal use.
OpenAI describes the four pilots:
- Apollo Research will assess whether models could secretly work against their developers' intentions, and whether safeguards would catch it.
- Goodfire will place researchers inside OpenAI to study model internals. Its findings are mainly meant for OpenAI researchers, and Goodfire will also publish learnings.
- METR will look for misaligned behavior in OpenAI's largest reinforcement learning run to date. That is a training method where a model learns from rewards. It will also look at a second review of OpenAI's monitors, which is still waiting on agreement about its scope.
- Transluce will examine the quality and coverage of OpenAI's post-training alignment tests. A longer follow-up of about eight weeks is planned for early 2027.
OpenAI also lists rules. Assessors who need sensitive information work on site, on OpenAI devices. OpenAI's Safety and Security Committee, an independent oversight group of the OpenAI Foundation's Board, will get unredacted reports. Unresolved disagreements go to that committee.
Assessors can publish their work independently. OpenAI says it keeps the right to remove confidential details and may ask for delays so that risks can be fixed first. Staff conversations are off the record, but assessors can ask OpenAI to confirm statements.
OpenAI discloses that Apollo Research and Goodfire are paid, and that the pay does not depend on results. METR and Transluce do not take payment but may accept computing credits.
What was done about it
Nothing is described as a response to a problem. OpenAI says it built these rules with input from the assessors and is sharing them to help set norms. It says the rules will likely change, and that it expects technical access problems and other surprises early on. It also says it is exploring future reviews of its compliance with its safety frameworks and of its organizational culture.
Why it matters
OpenAI says public reports from independent assessments can give people a clearer view of how labs evaluate and manage risks. It says the work may help build shared standards across the industry. The post does not yet include any findings. OpenAI notes that METR's scope is narrower than what METR proposed. Details of the Monitor assessment are still to be agreed. Read more in the original report.
This is a summary in everyday words, written by AI Misalignment Watch from the source above. If you spot a mistake, the original always wins.