OpenAI said on Nov. 16 that it has begun the first phase of what it calls embedded assessments, placing independent evaluators inside the company rather than handing them a finished model to poke at. Four organizations are involved: Apollo Research, Goodfire, METR and Transluce. The access list is the news. Depending on the evaluation, assessors can obtain model checkpoints for released and unreleased models, source code, training and evaluation data, monitoring systems and flagged behavior, technical documentation, and discussions with the researchers and engineers who did the work.
When direct access would expose the most sensitive intellectual property, such as weights or architecture, OpenAI says assessors can instead be paired with its own researchers to run analyses. Those who need to see classified material work on site, on OpenAI-provided devices, under security standards OpenAI says are at least as strict as those applied to employees.
Apollo Research is testing subversion: whether a model will quietly pursue goals other than the developer's, and whether OpenAI's safeguards can catch it. The work is split into two phases covering four claims, that training did not incentivize subversion, that strong evaluations and red-teaming failed to elicit it, that the model did not attempt it during internal deployment, and that the reasoning behind any subversion is detectable.
METR has two evaluations. The first is an alignment assessment of what OpenAI describes as its largest reinforcement learning run to date, searching for misalignment that existing tests and monitoring might miss, with METR able to set its own priorities and iterate on search, analysis and elicitation experiments. OpenAI's own footnote concedes that this scope is narrower than the broad misalignment investigation METR proposed. METR has already completed a three-week monitoring red-team exercise with access to OpenAI's codebase, monitoring systems and trajectories. A second, broader monitoring evaluation is agreed in principle, pending scope.
The other two cover different ground. Goodfire embeds interpretability researchers who use its platform to analyze internal mechanisms, aimed mainly at monitoring, debugging and aligning training runs, with findings published as the work matures. Transluce reviews the quality and coverage of OpenAI's post-training alignment evaluations, with both sides aiming at a follow-up of roughly eight weeks in early 2027 that would add broader model access and information on how evaluation results feed release decisions.
Oversight sits with the Safety and Security Committee, the independent committee of the OpenAI Foundation board, which receives unredacted reports and can be contacted privately by assessors during an evaluation. Disagreements escalate first to a safety advisory group and executives, and if unresolved go to the committee for a ruling. Staff interviews are off the record but come with an attestation pathway: assessors can ask OpenAI to certify a statement, and can publish either the certified statement or the fact that certification was refused. Apollo Research and Goodfire are paid, with payment not tied to findings; METR and Transluce take no fee but accept API credits or compute. All four must disclose conflicts and observe cooling-off periods before moving to another lab.
Two caveats keep this short of full independence. OpenAI retains the right to redact confidential information and may request responsible disclosure delays to fix problems before findings become public, and it acknowledges that early embedded work will hit "technical access challenges, space constraints, evolving security protocols, and other unexpected issues." The program also discharges a commitment OpenAI made under the White House accord on superintelligence to subject its safety controls to independent evaluation, and follows earlier pilots on internal monitoring and the investigation into the Hugging Face incident.
The precedent matters more than the paperwork. Anthropic floated the idea of embedding outside evaluators; OpenAI is the first lab to actually do it at this depth, which converts a long-running critique of self-assessment into a testable arrangement. The test will not be whether reports get published but whether they can say something a lab would rather not read, and whether access survives the first finding that does.
Comments (0)
Log in to join the discussion
Log InNo comments yet