OpenAI says it has signed agreements with four independent assessors, Apollo Research, METR, Transluce and Goodfire, for its “first phase of embedded assessments”. Its announcement was posted on X at 16:03 UTC on 16 November and is dated 13 November on its site. OpenAI says assessors may get “some of the same internal resources our researchers use”, including model checkpoints, source code, training and evaluation data and monitoring systems.
Apollo Research says six of its evaluators will work within OpenAI for four months, about 17 weeks by our arithmetic. Transluce says its pilot runs six weeks. OpenAI’s own footnote says two of the four, Apollo Research and Goodfire, accept financial compensation, while METR and Transluce do not.
Business Pill · The AI Audit Problem
A short Business Engineer explainer on the AI audit problem: when producing work becomes cheaper than checking it. It teaches the general idea only and does not discuss OpenAI or these assessments.
The key insight: As we read the terms, OpenAI is turning outside scrutiny into a written contract: four labs, six kinds of internal resource on the list, and a Safety and Security Committee that receives unredacted reports, while OpenAI keeps the right to redact confidential information from what the public sees.
What OpenAI Announced
OpenAI’s post says embedded assessments “give independent experts deep access to a lab’s internal models and systems to assess model alignment and the effectiveness of safeguards over an extended period of time”. It lists six kinds of resource: model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the researchers and engineers responsible for the work, by our count of its list.
The company says the first assessments “focus on alignment and safeguards for models during training and internal use”. It ties the program to a government commitment: “These efforts also support our commitments under the White House Accord on Super Intelligence, where we agreed to independent assessments of our safety controls”.
On X, OpenAI described the program as “giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards”. Apollo Research posted its own announcement three minutes earlier, at 16:00 UTC, according to the posts’ timestamps.

What Was Posted on X
METR, posting at 16:08 UTC on 16 November, says it has agreed with OpenAI to conduct a limited alignment assessment of OpenAI’s largest RL run and will publish one or more reports. The piece uses it for METR’s scope and publication plan.
Goodfire, posting at 16:14 UTC on 16 November, links its article announcing an embedded assessment with OpenAI, in which it says its researchers will use its interpretability platform to monitor training runs. The piece uses it for Goodfire’s scope.
What Each Lab Will Look At
Apollo Research, OpenAI says, will assess “whether models could covertly work against their developers’ intentions and whether our safeguards would detect such behavior”. OpenAI lists four claims Apollo will test, starting with “Training did not incentivize subversion” and ending with “Reasoning behind subversion would be detectable”. Apollo’s own post says it will publish two reports.
For METR, OpenAI says the first phase will assess “misaligned behavior in our largest reinforcement learning run to date, including failures that our existing tests and monitors may have missed”. METR posted that it “will publish one or more reports to share our findings and transparently describe our process”. A second METR monitor assessment, OpenAI says, is “pending mutual agreement on the assessment scope”.
OpenAI says Transluce will assess “our post-training alignment evaluations, examining their quality and coverage”. Transluce’s post says its staff will have access to “at least one internal-only model and the ability to review raw chain-of-thought outputs”, starting in early December. OpenAI says the two share the intent to run a follow-on assessment of “approximately eight weeks in early 2027”.
Goodfire, OpenAI says, will “embed interpretability researchers and use the Goodfire platform to analyze model internals”. Goodfire’s announcement calls the focus “model forensics” and says its researchers “will have deep access to OpenAI’s internal models and systems”.
How Much Access, Lab by Lab
OpenAI’s description of each engagement lists different resources. By our count across eight categories, it names six for Apollo Research, four each for METR and Transluce, and one for Goodfire, whose engagement it describes through the Goodfire platform rather than a list. A blank in the chart means OpenAI’s post does not list that item for that lab, not that access was refused.
OpenAI also sets a limit on the most sensitive material. Its post says that “When direct access could expose our most sensitive IP, such as model weights or architecture, we will consider alternatives”, including pairing assessors with OpenAI researchers to run analyses.

The Limits Written Into the Terms
Both OpenAI and the labs describe a right to redact. OpenAI says “we will retain the right to redact confidential information”. Apollo says OpenAI “may not redact Apollo Research’s verdicts for each claim”, but “remains empowered to redact evidence that they consider to reveal sensitive IP”.
Transluce says its agreement “prevents us from publishing results related to internal-only models”, and that OpenAI “must review and approve the embedded Transluce staff members prior to the pilot”. OpenAI says talks with employees are “off the record”, with a separate attestation pathway for statements assessors want to publish.
OpenAI says its Safety and Security Committee, an independent oversight committee of the OpenAI Foundation’s board, “will receive unredacted reports”. Apollo says the pilot “doesn’t yet fulfill the criteria above” for full embedded evaluations, while it meets Apollo’s bar for meta-transparency, the right to describe the limits of its own work.
Who Is Paid
OpenAI’s footnote says “Apollo Research and Goodfire accept financial compensation for their work, but their compensation is not contingent on the results of their assessment”. It says METR and Transluce do not accept payment but “may accept API credits / compute to use within their assessments”. Transluce says OpenAI will cover the cost of compute, provide on-site devices and make temporary workspace available.
On the same day, The Information reported that OpenAI had explored acquiring Goodfire at a price of more than $5 billion, as we reported. That report did not say a deal was agreed, and OpenAI’s post does not mention it.
Voluntary, Not Required
OpenAI presents the program as its own choice, tied to the White House Accord on Super Intelligence. In our earlier count, none of six federal and state measures we read requires an AI developer to get government approval before releasing a model, and the closest one is voluntary.
The Structural Read
As we read the four posts, the scope differs sharply by lab: OpenAI lists six access items for Apollo Research and one for Goodfire, by our count, and Transluce says it cannot publish results related to internal-only models.
As we read OpenAI’s footnote, the disclosure itself is the governance step: it names which two labs are paid and says their compensation is not contingent on the results of their assessment.
Apollo Research, 16 November 2026
“In general, the results of embedded evaluations must be credible independently of whether people trust the developer being evaluated.”
Three Implications
FOR AI BUYERS OpenAI says public reporting will help people understand the findings and their limitations; no findings have been published yet.
FOR OTHER LABS Apollo says it is working towards running full embedded evaluations with multiple AI developers in the future.
RELATED Our piece on the reported Goodfire talks covers the more-than-$5 billion price The Information reported.
The Business Engineer Lens
This story maps onto the Business Engineer framework Inside Anthropic’s Permission Layer.
The framework argues: “The governance layer sits at the intersection of capability, regulation, national security, access control, and geopolitical alignment.”
As we read it, OpenAI’s post is a set of access-control terms pointed inward: who outside the company may see checkpoints, code and monitoring, and under what redaction rules. The framework predates this program and does not discuss it.
What Is Not Established
No findings have been published. The four labs describe reports to come, and OpenAI says it may request “responsible disclosure delays” before publication. METR’s second assessment is pending agreement on scope, per OpenAI.
OpenAI’s post does not name the models or the training run under review, and it does not link the program to the timing of any model release. We did not contact OpenAI or the four labs for this piece.
The Bottom Line
OpenAI says four outside labs will get deep access to its internal models and systems, with six Apollo Research evaluators inside for four months and a six-week Transluce pilot from early December. Its own footnote says two of the four, Apollo Research and Goodfire, are paid, and its Safety and Security Committee will receive unredacted reports.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. The program, the four assessors, the access list, each lab’s scope, the redaction and oversight terms and the compensation footnote are from OpenAI’s announcement page, which is dated 13 November and which OpenAI posted on X at 16:03 UTC on 16 November. The six evaluators and four months, the two reports and the redaction limits on Apollo’s verdicts are from Apollo Research’s 16 November post; the six-week pilot, early-December start, internal-only model and staff-approval terms are from Transluce’s post dated 13 November; METR’s and Goodfire’s statements are from their own posts on X.
The Goodfire acquisition talks are as reported by The Information on 16 November. The counts of access items, the weeks for four months and the ratios are our arithmetic. OpenAI and the four labs were not contacted for this piece.
Sources: OpenAI, Our first phase of embedded assessments · Apollo Research, 16 Nov 2026 · Transluce, 13 Nov 2026 · Goodfire on X, 16 Nov 2026 · METR on X, 16 Nov 2026 · OpenAI on X, 16 Nov 2026









