OpenAI embedded assessments (Apollo, METR, Transluce, Goodfire)
Confirmed
Importance: major (4 of 5)The takeaway
On Nov 16, 2026 OpenAI announced a first phase of “embedded assessments”: Apollo Research, METR, Transluce and Goodfire get deeper access to internal models and systems to study alignment and test safeguards.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 18 November 2026
Your AI and this story
- GPT-6.2 Astra169 days after its cutoff
- Claude Opus 5.5139 days after its cutoff
- Gemini 4 Hydrogen230 days after its cutoff
- Grok 4.7169 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 139 days before it.
Key facts
- OpenAI on X (~880k views): ‘We’re working with @ApolloResearch, @METR_Evals, @TransluceAI, and @GoodfireAI on embedded assessments: giving outside experts deep access to our internal models and systems to investigate alignment and test our safeguards’
- METR on X (~198k views): ‘We have reached an agreement with OpenAI to conduct a limited alignment assessment. This pilot will focus on investigating the alignment properties of OpenAI’s largest RL run’; it will publish one or more reports; unpaid (may accept compute or API credits)
- Apollo Research (TechCrunch): studies whether training incentivises subversion of alignment or control, whether evaluations elicit it, whether it occurs in internal deployment and whether the reasoning is detectable; access to training code and datasets, RL rollouts with reasoning traces and rewards, and intermediate checkpoints, limited to agreed claims, models and periods; paid; starts this week
- Limits (TechCrunch, Apollo’s description): OpenAI may withhold information for technical infeasibility, security or IP, but agreed not to withhold it ‘just because it is unfavorable’; Apollo may talk only to ‘authorized staff’ selected by OpenAI; Apollo may publish its conclusions but OpenAI can limit supporting evidence; ‘We don’t get access to the general communication channels (e.g. Slack) an OpenAI safety researcher would’
- Transluce: whether OpenAI’s post-training safety evaluations are adequate, in 2027 how safety tests feed release decisions (unpaid). Goodfire: embedded interpretability researchers helping monitor training runs (paid)
- Governance: OpenAI’s Safety and Security Committee receives unredacted reports and can talk privately with evaluators; OpenAI calls the pilot ‘an ambitious first step towards deeper access and disclosure rights’ and ties it to its White House Accord commitments
- Critics: Buck Shlegeris (Redwood) said the agreements fall short of ‘employee-like access’, ‘disappointing but not surprising’ (TechCrunch); Steven Adler on X: ‘is this the “employee-like access” that Sam committed to back in September??’; Nathan Calvin (Encode) asked whether evaluators with this access would have detected the pre-Hugging-Face Artifactory outage, ‘I think the answer is no’
- Contrast: Anthropic gave six METR evaluators employee-like access with publication rights it cannot veto in October (Anthropic gives METR employee-like access and veto-free publication rights). METR and Redwood wrote the Hugging Face post-mortem; OpenAI earlier dismissed three researchers over sharing information with METR
Show 2 more
- Same day: California’s AG sued OpenAI over its agent-testing practices (California AG sues OpenAI over rogue agents)
- Reactions (Nov 16-18): Neel Nanda said the assessments are welcome but ‘fall short of Sam’s commitment to employee-like access’, since OpenAI controls the information flow; Daniel Kokotajlo (61k views) said OpenAI researchers tell him privately there will still be no thorough investigation of the Hugging Face incident; Mother Jones headlined that the ‘independent assessment’ does not look independent
What happened
OpenAI had said in September it would give independent evaluators access comparable to employees and announce contracts “in the coming weeks”. The Nov 16 post names four organisations with different scopes and payment terms. The openai.com page returned 403 to our fetcher; details here come from OpenAI’s and METR’s X posts and TechCrunch’s reading of the agreements.
Why it matters
Outside access to training runs, reasoning traces and checkpoints is what the Hugging Face post-mortem and the safety researchers’ letters asked for. Whether evaluators can find problems the lab does not point them to depends on the limits, which critics say leave OpenAI in control of whom evaluators talk to and what evidence becomes public.
Sources
7 sources from 3 sites. Numbers match the chips in the text.
7 sources: 3 primary, 1 press, 3 reactions
Primary
- OpenAI: Our first phase of embedded assessments (Nov 16)openai.com, official
- OpenAI on Xx.com, official
- METR on Xx.com, official
Press
- TechCrunch: OpenAI will embed safety researchers to investigate AI subversiontechcrunch.com, press
Reactions
- Steven Adler on Xx.com, discussion
- Nathan Calvin on Xx.com, discussion
- Daniel Kokotajlo on X: still no thorough investigation of the HF incidentx.com, discussion
Changes
- Nanda, Kokotajlo and Mother Jones reactions
- Filed (news agent; X via fxtwitter)