OpenAI signs four embedded assessors, and Apollo says six researchers will spend four months inside, without Slack
OpenAI’s post is dated 16 November. Ashley Capoot’s CNBC story is stamped 12:54 p.m. Eastern and updated at 4:54 p.m. Apollo’s own post the same day says the pilot does not yet meet the access and disclosure bar it wants, and that verdicts cannot be redacted even when the evidence under them can.
Feature4 minIn-depth
James Okonkwo4 min read
- OpenAI
- Apollo Research
- METR
- Evaluations

Photo: OpenAISource page
OpenAI said on Monday that it has signed agreements with Apollo Research, METR, Transluce, and Goodfire for a first phase of embedded assessments. The post, dated 16 November, describes the arrangement as giving independent experts access, over an extended period, to internal models and systems, including checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and the researchers responsible for the work. The initial assessments are about alignment and safeguards during training and internal use. Ashley Capoot’s CNBC story is stamped 12:54 p.m. Eastern and updated at 4:54 p.m. She writes that Sam Altman pledged “employee-like access” for independent evaluators in September, after Anthropic’s Dario Amodei said his company would do the same.
The four scopes are not the same job. Apollo will assess whether models could covertly work against their developers’ intentions and whether safeguards would detect that, against four claims: training did not incentivize subversion, strong evaluations and red teaming do not elicit it, models have not attempted it during internal use, and reasoning behind it would be detectable. Goodfire will embed interpretability researchers and use its platform on training-run monitoring, debugging, and alignment. OpenAI says the primary audience for Goodfire’s assessments is internal researchers, and that Goodfire will also publish as the work goes. METR is down for two assessments OpenAI says it intends to conduct. The first looks for misaligned behavior in the largest reinforcement-learning run to date, including failures existing tests may have missed. A footnote says that scope is less expansive than the broad survey in METR’s own proposal. A second, broader monitor assessment is “pending mutual agreement.” Transluce will review the quality and coverage of post-training alignment evaluations. OpenAI and Transluce also share an intent for a follow-on of about eight weeks in early 2027, with broader model access and information on how evaluation results inform release decisions.
Six people, four months, and no general Slack
Apollo’s own post, also dated 16 November, is more specific about the people and the limits. A team of six will work inside OpenAI for four months, assessing alignment through training, evaluations, and internal deployment under scoped rules of access. The goal it states is to investigate whether models attempt to subvert their own alignment or control mechanisms, or those of their successors. For the training claim, access includes relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate checkpoints. Access for the other claims, Apollo writes, remains to be agreed. All of the work will be done on OpenAI hardware. Evaluators have separate channels that can be monitored. “By default, evaluators do not have access to any of the general channels (e.g. Slack) that a safety researcher at OpenAI would.” If an employee outside the approved group raises confidential information, Apollo must stop the substantive discussion and route follow-up through a designated contact.
OpenAI has agreed not to withhold information or access to staff, nor request redactions, simply because the information or findings are unfavorable.
— Apollo Research, 16 November 2026
That sentence is the bar Apollo says this pilot meets, and it is not the bar Apollo says a full embedded evaluation would meet. OpenAI may still withhold in-scope items for technical infeasibility, security, or intellectual property, and must say that withheld items exist and give written reasons. Where it considers it feasible, it will identify withheld records by category and approximate volume. Expansion of scope is not guaranteed. Apollo will publish two reports. The first covers training incentives for subversion. The second covers evaluations, internal deployment, and whether subversive reasoning would be detectable. OpenAI may not redact the verdicts. It may redact evidence it considers sensitive intellectual property, including evidence that supports a finding. Readers of the public report may not see all of the evidence under a conclusion. Apollo says it will be able to indicate whether redactions affected the conclusions, and to say when an attestation was declined. OpenAI will share the unredacted reports with the board’s Safety and Security Committee.
Who is paid, and who may delay a sentence
OpenAI’s footnote on money is explicit. Apollo and Goodfire accept financial compensation that is not contingent on the results. METR and Transluce do not accept financial compensation, and may accept API credits or compute for use inside the assessment. “We do not believe that accepting financial compensation prevents assessors from maintaining independence – but we believe it is important to disclose this,” the company writes. Capoot’s story prints the same split. Staff discussions that inform an assessment are off the record and cannot be published or attributed. A separate attestation can verify, correct, or clarify something learned in conversation, without naming the employee. Assessors may publish a statement OpenAI attests to, or disclose that it declined. Disagreements go to the Safety Advisory Group and executives, the committee is notified, and an unresolved dispute goes to the committee. OpenAI also keeps a right to redact confidential information and to request delays so it can act on findings it says pose security or safety risks. It asks participating staff at the outside firms to observe a brief cool-off before embedding at another lab.
The post ties the agreements to commitments under the White House Accord on Super Intelligence, which it says included independent assessments of safety controls. It also says the company expects technical access problems, space constraints, and changing security rules in the early stages, and that it is exploring later assessments of compliance with its safety frameworks and of organizational culture. Apollo’s last line on the design is that without the right to explain what was withheld, a lack of reported problems can look reassuring when it only reflects what an evaluator was allowed to see. The four-month clock, and the two reports, are the parts of that argument with a date attached. The post does not give the day the six people start.


