Four outside groups get deep access to OpenAI's models to check its safety work
OpenAI says it has signed agreements with four independent groups for "embedded assessments", in which outside experts use some of the same internal resources as its own researchers to check its models.
OpenAI says outside experts will get deep access to its internal models, and some will work inside its offices, in a post dated 13 November 2026.
The company says it has signed agreements with four independent assessors: Apollo Research, METR, Transluce and Goodfire.
OpenAI defines "embedded assessments" as giving independent experts deep access to a lab's internal models and systems, over an extended period, to assess model alignment and how well safeguards work. Alignment means whether a model acts as its developers intend.
OpenAI says assessors get some of the same resources its researchers use: source code, training data, monitoring systems and talks with the staff responsible. The company says this lets them find failures it may not have anticipated.
According to OpenAI, Apollo Research will assess whether models could secretly work against their developers' intentions, and whether OpenAI's safeguards would detect it.
Goodfire will embed interpretability researchers, who study what happens inside a model, and use its own platform to analyse model internals. Goodfire will also publish findings as the work goes on.
METR will look for misaligned behaviour in OpenAI's largest run of reinforcement learning (a training method in which a model is rewarded for desired results) to date, OpenAI says. OpenAI says this scope is narrower than what METR itself outlines.
A broader assessment of its monitoring systems is agreed only preliminarily, pending agreement on its scope, OpenAI says.
Transluce will check the quality and coverage of OpenAI's alignment tests run after training. OpenAI and Transluce say they intend a follow-on assessment of about eight weeks in early 2027.
OpenAI says assessors who need sensitive confidential information will work on site, on OpenAI devices, mainly to protect that information. The company says its Safety and Security Committee, which it describes as an independent oversight committee of the OpenAI Foundation's Board, will receive unredacted reports and can speak with assessors, including privately.
OpenAI says assessors keep independence in publishing, but it keeps the right to redact confidential information, and may ask for publication to be delayed to fix safety or security issues.
Apollo Research and Goodfire are paid, OpenAI says, though not according to their results. METR and Transluce accept no payment, but may accept computing credits.
OpenAI says it expects early difficulties, including technical access problems, limited space and changing security rules.
OpenAI says the work supports its commitments under the White House Accord on Super Intelligence, where it agreed to independent assessments of its safety controls.
OpenAI says it is also exploring future assessments of compliance with its safety frameworks, and of its organisational culture and processes.
Sources
- Our first phase of embedded assessments · openai.com