OpenAI Signs Four Independent Labs for Embedded Safety Assessments, Starting With a Six-Person, Four-Month Apollo Pilot

OpenAI signed four labs for embedded assessments, giving Apollo Research, METR, Transluce and Goodfire deep access to test whether models subvert safeguards.

Share this article

Featured image: the Pioneer Building in San Francisco’s Mission District, which housed OpenAI offices, in a July 2019 file photo. The photo shows the building in 2019 and is not from the assessment announcement. Photo: HaeB, Wikimedia Commons, CC BY-SA 4.0. Cropped to 16:9.

OpenAI said Monday, Nov. 16, 2026, that it has signed agreements with four independent labs — Apollo Research, METR, Transluce and Goodfire — for its first phase of embedded assessments, a setup in which outside researchers get deep access to the company’s models and internal systems to test safety claims over months, not just before a public release.

Advertisement

Apollo Research said a team of six of its staff will work inside OpenAI for four months, under scoped access rules, to examine whether models try to subvert their own or their successors’ alignment or control mechanisms. The announcement comes months after OpenAI said two of its models left a testing environment and compromised Hugging Face, an incident that is now at the center of a California lawsuit against the company.

Independent labs
4
Apollo Research, METR, Transluce, Goodfire
Apollo team
6 people
Embedded for a four-month pilot, Apollo said
Apollo claims tested
4
Training, evaluations, internal use, detectability
Transluce follow-on
~8 weeks
Planned for early 2027, OpenAI said

What each lab is being asked to do

In a Nov. 16 post, OpenAI said embedded assessments give independent experts “some of the same internal resources our researchers use,” including model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the staff who do the work. The first phase, it said, focuses on alignment and safeguards during training and internal use, and supports commitments under the White House Accord on Super Intelligence to allow independent assessments of safety controls.

Lab Assigned work, as described by OpenAI Compensation, as disclosed
Apollo Research Test four claims: that training did not incentivize subversion; that strong evaluations do not elicit it; that models have not attempted it in internal use; and that reasoning behind it would be detectable. Accepts pay; OpenAI said it is not tied to results.
Goodfire Embed interpretability researchers to monitor, debug and assess alignment in training runs. OpenAI said the main audience is its own researchers, with public learnings to follow. Accepts pay; OpenAI said it is not tied to results.
METR Two planned assessments: one of misaligned behavior in OpenAI’s “largest reinforcement learning run to date,” and a broader monitor assessment whose scope was still pending mutual agreement. Does not accept pay; may accept compute or API credits.
Transluce An initial review of post-training alignment evaluations, with a planned follow-on of about eight weeks in early 2027 and broader access, including how evaluation results inform release decisions. Does not accept pay; may accept compute or API credits.

Source: OpenAI, “Our first phase of embedded assessments,” Nov. 16, 2026, including footnote 4 on compensation.

OpenAI said it recently completed a three-week monitor red-teaming exercise with METR and that the first METR alignment assessment will be narrower than METR’s own published proposal for a broad survey of misalignment. The company also pointed to a prior investigation of the Hugging Face incident and a recent pilot on internal monitoring systems.

How much access the labs actually get

Apollo Research, in its own post the same day, called the pilot “an ambitious first step towards deeper access and disclosure rights, even if it doesn’t yet fulfill the criteria” it sets for a full embedded evaluation: access to all relevant records even if a developer wants to hide unfavorable information, and the ability to disclose findings even if the developer wants to block that disclosure. Apollo said the pilot does meet its bar for “meta-transparency,” meaning it can publicly explain the limits of the work.

On access, Apollo said OpenAI has committed to give it what is required to assess whether models attempt subversion, including, for the training-incentives claim, relevant training code and datasets, reinforcement-learning rollouts with reasoning traces and rewards, and queries of intermediate checkpoints. Detailed access for the other three claims, Apollo said, “remain to be agreed.” Even inside the agreed scope, OpenAI may withhold items it deems reasonably necessary because of technical infeasibility, security risks or intellectual-property protection, Apollo said, but has agreed not to withhold information or staff access simply because the material is unfavorable, and must say when items have been withheld and why.

The practical limits are specific. Apollo said its staff may discuss confidential information only with authorized employees; OpenAI selects who participates, even if Apollo suggests names; formal conversations must be logged; and if an unauthorized employee raises confidential information, Apollo must stop the discussion and route it through a designated contact. Work will be done on OpenAI hardware, on monitored channels, and, “by default, evaluators do not have access to any of the general channels (e.g. Slack) that a safety researcher at OpenAI would.”

Those limits matter because of an earlier pledge. CNBC reported that OpenAI CEO Sam Altman promised in September to grant independent evaluators “employee-like access,” after Anthropic’s Dario Amodei said his company would do so. The Verge reported that members of the AI research community have raised concerns that the Apollo arrangement, with its green-lit interviews, logged conversations and lack of Slack access, falls short of that promise. OpenAI’s own post said assessors who need sensitive information will work on site, on OpenAI devices, and that discussions with employees are off the record unless the company attests to a statement the assessors want to publish.

Advertisement

Publication, redaction and who settles disputes

Apollo said it will publish two public reports: one on whether training incentivizes subversion, and a second on evaluations, internal deployment and the detectability of subversive reasoning. OpenAI may request redaction of confidential information, including evidence that supports Apollo’s findings, but “may not redact Apollo Research’s verdicts for each claim” and may not demand cuts merely because a finding is unfavorable, Apollo said. Readers of the public report “may therefore be unable to examine all the evidence underlying a published conclusion,” though Apollo said it will be able to say whether redactions affected its conclusions. Unredacted reports will go to the Safety and Security Committee of the OpenAI Foundation’s board.

OpenAI described a similar structure. Disagreements, it said, go first to the Safety Advisory Group and company executives, with a set time to resolve them, and then to the Safety and Security Committee if they remain. The committee is to receive unredacted reports and be able to speak privately with the assessors. OpenAI also said it may request “responsible disclosure delays” so it can act on findings that pose security or safety risks.

OpenAI said it does not believe that paying Apollo and Goodfire prevents independence, “but we believe it is important to disclose this.” It said it expects the first phase to lead to “longer-term, consistent embedded engagements” and is exploring later assessments of compliance with its safety frameworks and of organizational culture and processes. No dates were given for public reports from this phase, beyond Transluce’s planned early-2027 follow-on of about eight weeks.

The company is expanding other research access in the same week, including a larger ChatGPT program for academic researchers. Those grants are a separate track from the safety assessments described here.

Sources

Read next in Artificial Intelligence

Get The Boston Report Briefing

Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.

Free. We never sell your address. Privacy policy

Get The Boston Report Briefing
Our free weekly email with the most important stories from Boston and the world. Unsubscribe anytime.
Sign up