What OpenAI Signed On 16 November
OpenAI announced on 16 November that it signed Apollo Research, METR, Transluce and Goodfire for the first phase of embedded assessments. Access can include model checkpoints, source code, training data and monitoring systems. Apollo tests whether training rewarded subversion and whether it would be detectable, METR looks at what OpenAI calls its largest reinforcement-learning run to date, Transluce audits the quality and coverage of its post-training alignment evaluations, and Goodfire embeds interpretability researchers. Apollo and Goodfire accept pay that does not depend on results, while METR and Transluce take no money, though they may accept API credits or compute.
Where The Terms Limit Independence
The assessors keep the right to publish, but OpenAI can redact confidential information and request responsible-disclosure delays. Disputes go first to its Safety Advisory Group and executives, and only then to the Safety and Security Committee, which also receives unredacted reports. OpenAI can decline to attest to statements from off-record employee conversations, and assessors may then say only that it declined. Conflict-disclosure rules are still being written, METR's monitor assessment has no agreed scope, and the eight-week follow-on planned for early 2027 is an intention, not a commitment.
What To Ask A Vendor That Cites Outside Testing
In our view, embedded access is a real step beyond an audit of a few days, but it is assurance only to the extent the terms are known. If an AI vendor says outsiders assessed its models, ask what the assessor could see, who can redact the report, who settles disputes and whether money changed hands. Keep the answers in your procurement file, because the same questions apply to every lab, European ones included.
Read next: The Independent Audit That Needed the Vendor's OK | Outside Researchers Caught Three of Four OpenAI Hacks



