TL;DR
OpenAI said on 16 November that four outside organisations, Goodfire, Transluce, METR and Apollo Research, have signed up to work inside the company examining its models’ alignment and safeguards. Assessors will get access to unreleased models, training checkpoints, training and evaluation data and monitoring systems, working on site on OpenAI devices. The company has published the protocols it agreed with them.
OpenAI calls the arrangement embedded assessment: outside experts with the kind of internal access its own researchers have, over an extended period. It links the work to promises it made in the White House superintelligence accord, which include outside checks on its safety controls.
What each assessor will examine
Apollo Research will look for signs that models might secretly act against what their makers intend, and test whether OpenAI’s safeguards would spot it. It will assess four claims, among them that training did not reward subversion and that any reasoning behind it would be detectable.
METR plans two pieces of work. The first examines misaligned behaviour in what OpenAI describes as its largest reinforcement learning run so far; a second, broader review of OpenAI’s monitors is agreed in principle but still being scoped. Transluce will review the quality and coverage of OpenAI’s post-training alignment evaluations, with a longer follow-up of about eight weeks pencilled in for early 2027. Goodfire will place interpretability researchers inside the company to study model internals during training.
The rules of engagement
Disputes between OpenAI and an assessor go first to its Safety Advisory Group and executives, then to the Safety and Security Committee, an independent body drawn from the board of the OpenAI Foundation, which also receives unredacted reports. Staff interviews are off the record, but assessors can ask OpenAI to formally attest to statements they want to publish, or say publicly that it declined.
Assessors keep editorial independence, though OpenAI can redact confidential material and request delays to fix safety or security issues first. Apollo and Goodfire are paid; METR and Transluce take no fee, though they may use API credits or compute provided for the work.
Looking forward
The UK has its own stake in this model of scrutiny. Kanishka Narayan, the UK’s AI minister, told Axios in an interview published on 16 November that the government’s approach rests heavily on the AI Security Institute and that he wants its model testing seen as a resource for Americans as well as Britons. In our view, OpenAI publishing its access terms gives regulators and institutes like AISI a concrete benchmark: what deep access looks like in practice, and where a lab still holds back, such as model weights and architecture, which OpenAI says it may shield by pairing assessors with its own researchers.