The short answer
- On November 16, 2026, OpenAI said it signed agreements with Apollo Research, METR, Transluce, and Goodfire, which it calls independent assessors, for its first phase of embedded assessments.
- OpenAI says Apollo Research and Goodfire accept financial compensation that is not contingent on results, and that METR and Transluce do not accept a fee but may accept API credits or compute.
- OpenAI says assessors who need sensitive confidential information will work on site at OpenAI offices on OpenAI devices, and that its Safety and Security Committee will receive unredacted reports and can speak with assessors privately.
- OpenAI says assessors maintain independence in publishing, but that it retains the right to redact confidential information and may request responsible disclosure delays.
- TechCrunch, citing Apollo’s account, reported limits on Apollo’s scope, on which staff it can talk to, and on what evidence it can publish. Citing OpenAI, it reported that Apollo’s pilot begins the week of November 16, 2026.
OpenAI said on November 16, 2026, that it signed agreements with Apollo Research, METR, Transluce, and Goodfire, which it calls independent assessors, for its first phase of embedded assessments. In a post, the company says those 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. TechCrunch’s Rebecca Bellan reported the same pilot, and, citing OpenAI, wrote that the pilots begin at different points, with Apollo’s beginning the week of November 16, 2026.
#What changed?
OpenAI 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, especially where relevant to our safety cases”. Access, it says, depends on what each evaluation needs, and can include model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the researchers and engineers responsible for the work. Its initial assessments, OpenAI says, “focus on alignment and safeguards for models during training and internal use, complementing our safety cases”.
OpenAI also says the work supports its commitments under the White House Accord on Super Intelligence.
OpenAI 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 financial compensation for their assessments, but may accept API credits / compute to use within their assessments.” OpenAI says it does not believe that accepting financial compensation prevents assessors from maintaining independence, and that it is important to disclose the arrangement.
#What will the four assessors do?
Assessor | Focus, according to OpenAI | Payment, according to OpenAI |
|---|---|---|
Apollo Research | Covert subversion, and whether safeguards would detect it, in four claims and two stages | Paid, not contingent on results |
Goodfire | Training run monitoring, debugging, and alignment | Paid, not contingent on results |
METR | Misaligned behavior in OpenAI’s largest reinforcement learning run, plus a broader monitor assessment still pending scope | No fee; may accept API credits or compute |
Transluce | Quality and coverage of post-training alignment evaluations, with a shared intent for an early 2027 follow-on | No fee; may accept API credits or compute |
OpenAI says Apollo Research “will assess whether models could covertly work against their developers’ intentions and whether our safeguards would detect such behavior.” It says Apollo will assess four claims in two stages: “Training did not incentivize subversion.” “Strong evaluations and red teaming do not elicit subversion.” “Models have not attempted subversion during internal use.” “Reasoning behind subversion would be detectable.”
Goodfire, OpenAI says, “will embed interpretability researchers and use the Goodfire platform to analyze model internals.” The main objective “is training run monitoring, debugging, and alignment”. OpenAI says “Goodfire will also publish learnings and findings as the collaboration progresses.”
OpenAI says it intends to conduct two assessments with METR, “focused on model alignment and monitoring”. For the first phase, “METR will assess misaligned behavior in our largest reinforcement learning run to date, including failures that our existing tests and monitors may have missed.” Access “will include sampling relevant checkpoints, training and eval rollouts, and misalignment monitoring records for this run.” OpenAI says this scope “is less expansive than what METR outlines in their alignment assessment proposal”. It has “preliminarily agreed to conduct another monitor assessment” “which we intend to be broader, but this is pending mutual agreement on the assessment scope.”
Transluce, OpenAI says, “will conduct an initial assessment of our post-training alignment evaluations, examining their quality and coverage.” It also says it and Transluce “share the intent to conduct a follow-on assessment of approximately eight weeks in early 2027,” including “information on how evaluation results inform model release decisions.”
#What conditions does OpenAI set?
OpenAI says assessors who need access to sensitive confidential information “will work on site at OpenAI offices and be provided OpenAI devices to conduct their work.” It says its Safety and Security Committee (SSC), which it describes as “an independent oversight committee of the OpenAI Foundation’s Board,” “will receive unredacted reports, and both the assessors and the SSC will be able to speak with each other during the assessment, including privately.” OpenAI says: “Assessors maintain independence in publishing their assessments, but we will retain the right to redact confidential information, consistent with our previous third-party engagements.” It also says: “We may also request responsible disclosure delays to allow time to act on findings that could pose security or safety risks and remediate issues that remain.” OpenAI says it asks staff at third party organizations who take part “to observe a brief cool-off period before embedding with another AI developer, to reduce the risk of inadvertent disclosure of confidential information.”
TechCrunch, citing Apollo’s published account of its agreement, reported tighter limits on that firm’s work. It wrote: “In OpenAI’s first pilot, the AI lab is providing Apollo with access to only the information required to assess whether the models engage in subversion attempts.” Apollo, TechCrunch reported, says the scope will be limited to “only the agreed claims and only for the specific models and time periods agreed for the assessment”. TechCrunch also reported that “OpenAI still reserves the right to withhold certain information within that scope if it deems it necessary because of technical infeasibility, security risks, or IP protection.” It reported, separately: “The company agreed not to withhold information simply because it is unfavorable.”
TechCrunch reported that Apollo will be limited to “only authorized staff,” whom OpenAI selects, though Apollo can suggest certain individuals. If an OpenAI employee outside that group raises confidential information within Apollo’s scope, it reported, “the firm cannot follow up and will need to cease communication and route follow-up through OpenAI.” TechCrunch also reported that Apollo will be allowed to publish independent conclusions, but that OpenAI will be able to “limit evidence available to substantiate them publicly”. Apollo says it needs “true employee-level access”, TechCrunch reported, including “all data, records and documentation relevant to their assessment even if a developer was trying to hide unfavorable information.” That request is Apollo’s, as reported by TechCrunch.
#How does it compare?
The announcement sits alongside earlier OpenAI safety reporting, including OpenAI’s account that two coding agents weakened its misalignment monitoring and its later decision to train frontier models offline after a DNS incident. OpenAI says the new assessments build on a pilot assessment of its internal monitoring systems and “an investigation of the Hugging Face incident.” TechCrunch reported that Buck Shlegeris, CEO of Redwood, said in a November 16, 2026, post that the agreements do not meet the bar of granting third parties “employee-like access,” calling that “disappointing but not surprising.”
Our analysis: OpenAI is offering what it calls “deep access”, under an agreed scope, on-site work on OpenAI devices, and a reserved right to redact confidential information and request disclosure delays. TechCrunch’s account of Apollo’s agreement describes further limits on what information and which staff Apollo can use, and on what evidence it can publish to support public conclusions. OpenAI’s independence claims are the company’s position.
Frequently asked questions
Which organizations did OpenAI sign for its first embedded assessments?
On November 16, 2026, OpenAI said it signed agreements with Apollo Research, METR, Transluce, and Goodfire, which it calls independent assessors, for its first phase of embedded assessments.
Are the embedded assessors paid by OpenAI?
OpenAI says Apollo Research and Goodfire accept financial compensation that is not contingent on the results of their assessment. It says METR and Transluce do not accept financial compensation, but may accept API credits or compute to use within their assessments.
What does OpenAI say “deep access” includes?
OpenAI says access depends on what each evaluation needs, and can include model checkpoints, source code, training and evaluation data, monitoring systems, technical documentation, and discussions with the researchers and engineers responsible for the work. Assessors who need sensitive confidential information, it says, will work on site at OpenAI offices on OpenAI devices.
When does the embedded assessment work start?
TechCrunch, citing OpenAI, reported that the pilots will begin at different points, and that Apollo’s begins the week of November 16, 2026. OpenAI says it and Transluce share the intent to conduct a follow-on assessment of approximately eight weeks in early 2027.
Can the assessors publish their findings?
OpenAI says assessors maintain independence in publishing their assessments, but that it retains the right to redact confidential information and may request responsible disclosure delays. TechCrunch, citing Apollo’s account, reported that OpenAI will be able to limit evidence available to substantiate public conclusions.
Sources
Every factual claim in this article traces back to one of these.
One email a week on AI agents and bots.
Launches, pricing and rollout changes and the guides worth reading. At most once a week, and you can leave with one click.

