OpenAI to embed six Apollo Research evaluators for a four-month safety pilot
Apollo Research will work inside OpenAI during training and internal deployment to test whether models try to subvert their own safeguards, with published reports and stated limits.
· 7 min read
Written by AI from the sources linked below. How we use AI
On this page
OpenAI has agreed to let outside safety researchers work inside the company while its models are being trained, not only after they are built. Apollo Research said on 16 November 2026 that a team of six of its evaluators will spend four months at OpenAI in a pilot of what the field calls "embedded evaluations". The terms matter as much as the headline: both sides have published enough detail to show what the evaluators will, and will not, be able to see and say.
Key points
- Apollo Research will place a team of six inside OpenAI for four months, checking how models are aligned during training, evaluation and internal deployment, according to Apollo's announcement.
- The pilot asks whether models try to subvert their own or their successors' alignment or control mechanisms, across four specific questions listed below.
- OpenAI agreed not to withhold information, staff access or redactions simply because the findings are unfavourable, Apollo says, but it can still withhold in-scope items for technical infeasibility, security risks or IP protection.
- Apollo will publish two public reports, and OpenAI may request redaction of evidence it considers sensitive IP, though not of Apollo's verdicts on each claim.
- OpenAI says it has signed agreements with four independent assessors (Apollo Research, METR, Transluce and Goodfire) for its first phase of embedded assessments, according to a summary of its post; we could not open OpenAI's page.
What happened
On Monday 16 November, Apollo Research published "A Pilot Towards Embedded Evaluations". Six Apollo staff will work within OpenAI for four months, "assessing the alignment of models throughout training, evaluations, and internal deployment with scoped rules of access", in Apollo's words. Their goal is to investigate whether models attempt to subvert their own or their successors' alignment or control mechanisms.
Apollo lists four areas the pilot will examine across development:
- Whether training incentivizes subversion.
- Whether evaluations can elicit subversion.
- Whether subversion occurs during internal deployment.
- Whether reasoning relevant to subversion would be detectable.
TechCrunch's report of the same day says Apollo did not share when the pilot will begin, and that TechCrunch had asked OpenAI for comment. The same day, OpenAI published its own post, "Our first phase of embedded assessments". Its page would not load for us, so we rely on Techmeme's summary of its first line, which says OpenAI has signed agreements with four independent assessors (Apollo Research, METR, Transluce and Goodfire). We cannot say what the work of the other three involves.
Background
Until now, outside evaluators have typically tested a model after post-training and shortly before public release. Apollo says that approach "is not sufficient" for testing misaligned behaviours such as scheming, or for non-public uses of frontier models inside a developer during internal deployment. TechCrunch reports that the idea of embedding evaluators in frontier labs first came from Anthropic, and that OpenAI appears to be the first to act on it.
Apollo's blog says the pilot builds on earlier joint work with OpenAI: stress-testing OpenAI's training interventions designed to prevent scheming, showing how a model's awareness of evaluation and oversight affects its behaviour, and showing that models can increasingly favour grader preferences over user instructions as reinforcement learning progresses.
How the pilot works
Apollo frames what any credible embedded evaluation needs as three things. Access: all data, records and documentation relevant to the assessment, even if a developer was trying to hide unfavourable information. Disclosure: the right to share findings with an appropriate outside audience even if the developer tried to prevent it. Meta-transparency: the right to explain, in enough detail for outsiders, the constraints on access and disclosure, so readers know what a report can and cannot establish.
Apollo is open that this pilot does not yet meet the first two. It says the pilot "doesn't yet fulfill the criteria above", but that it does meet the bar for meta-transparency.
Access. For the question of whether training incentivises subversion, OpenAI has committed to give Apollo the relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints. Access arrangements for the other claims "remain to be agreed". The assessment covers only agreed claims, specific models and time periods. OpenAI may withhold in-scope items when it considers it reasonably necessary, but it has agreed not to withhold information or staff access simply because it is unfavourable, must tell Apollo that withheld items exist, and must give written reasons for refusals. Where feasible it will identify withheld records by category and approximate volume. Apollo can ask for more access and escalate refusals, but expansion is not guaranteed.
Staff. Apollo may discuss confidential OpenAI information only with authorised staff. It can suggest people, but OpenAI chooses the participants, and formal conversations must be logged. If an OpenAI employee outside that group raises confidential in-scope information, Apollo must stop substantive discussion and route follow-up through its designated OpenAI contact. All the work is done on OpenAI hardware, evaluators have separate channels that can be monitored, and by default they do not have access to general channels such as Slack.
Disclosure. Apollo will publish two reports. The first covers training incentives for subversion. The second covers evaluations, internal deployment and whether subversive reasoning could be detected. Each will set out methods, conclusions and limitations, including questions that could not be assessed. OpenAI will review them and may ask for confidential information to be redacted, including supporting evidence. It may not redact Apollo's verdict on each claim and may not request redactions merely because a finding is unfavourable. OpenAI will also share the unredacted reports with the Safety and Security Committee of its Board. Claims based on employee discussions can only be published with OpenAI's formal attestation, unless Apollo establishes them independently. Apollo can say publicly that an attestation was declined and explain how that affects its conclusions.
Meta-transparency. Apollo may describe the scope and conditions of the engagement, the access requested and received, what was withheld or delayed and why, whether material was removed from a report at OpenAI's request and how much that matters to its conclusions, and any pressure it believes shaped its work. It can discuss its access and publication terms publicly without OpenAI's prior review.
What it means
For people who follow AI safety, the interesting part is the structure rather than the announcement. A report from an evaluator who has declared what it was not allowed to see is more informative than one that says nothing. Apollo's own argument is that "without meta-transparency, a lack of reported problems can look reassuring even when it simply reflects limits on what an evaluator could investigate or disclose."
For developers and companies that buy or build on frontier models, the pilot is a signal about how much trust to place in lab safety claims. Evaluations done only before release can say little about what happens inside training runs. Evaluators inside the lab can in principle check claims about how a model was trained and how it behaved, but only within a scope the lab agrees to. The reports will tell readers how wide that scope was.
For labs, Apollo calls the agreement "a public benchmark for other evaluation engagements". Apollo says it is working towards full embedded evaluations with multiple developers, and that it hopes to do so with deeper access and stronger disclosure rights.
TechCrunch notes that some limits remain areas of concern for Apollo. It reports that Apollo believes a truly independent evaluation would need employee-level access to all relevant data, and that OpenAI's right to limit the evidence behind published conclusions could make it hard for outsiders to assess the findings. Apollo's own post is more measured, calling the pilot "an ambitious first step towards deeper access and disclosure rights".
What we don't know yet
- When the four months begin, and so when the two reports can be expected. Apollo has not said, according to TechCrunch.
- What OpenAI's own post says in full. We could not open it, and rely on Techmeme's summary of the first line for the four assessors.
- What METR, Transluce and Goodfire will do in this first phase, what terms they have agreed, and whether they get the same access and disclosure rights as Apollo.
- Which models and time periods are in scope. Apollo says they are "agreed", but does not name them in the post.
- How much of the reports will be redacted, and how much redactions will weaken the conclusions. Apollo says it can say so, which is the point of the meta-transparency terms.
What to watch next
Watch for Apollo's first report, on whether training incentivises subversion, and for how it describes any withheld material. Watch for whether the access arrangements for the three other claims get agreed, and published. And watch whether other developers sign similar terms: Apollo says it wants to run embedded evaluations with several, and has stated its minimum standard in public.
Sources
- Apollo Research: A Pilot Towards Embedded Evaluations (16 November 2026, primary source)
- TechCrunch: OpenAI will embed Apollo researchers to investigate AI subversion
- OpenAI: Our first phase of embedded assessments (page would not open for us; the list of four assessors is as summarised by Techmeme)
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
ChatGPT Wallet: OpenAI's dots agents can now check out with your card
OpenAI's new Wallet starts with some U.S. Pro subscribers; each purchase needs your approval and the agent never sees the card number.
audiotale.net team5 min read
METR: Anthropic's AI research is running about 2x faster, but most evidence is redacted
METR's outside assessment estimates Anthropic's capability progress at about twice its earlier pace, mostly due to AI, while Anthropic redacted most of the supporting evidence.
audiotale.net team8 min read
GPT-6 Cyber: OpenAI's Critical-rated exploit model scores 48.5% on ExploitGym, gated to Daybreak Red
OpenAI's new security model beats GPT-6 Astra on its own exploit benchmark and, the company says, stays in scope over 99% of the time, but access starts with approved red-team customers only.
audiotale.net team7 min read