Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.

Editorial · CASRAI · AI and ML research outputs

Apollo details four-month OpenAI safety pilot

Apollo will embed six evaluators at OpenAI for four months, with two public reports and explicit limits on access and disclosure.

Published 16 Nov 2026· 5 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

Apollo Research announced on November 16, 2026 that six of its evaluators will work within OpenAI for four months to investigate whether AI models undermine their own or successor models’ safeguards. The pilot opens parts of training and internal deployment to outside assessment, while detailed restrictions on access and publication shape what university researchers and other readers will be able to conclude.

What the new Apollo pilot will investigate

In its November 16 announcement, Apollo sets out four questions: whether training encourages subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether the reasoning behind it would be detectable. The assessment covers models’ attempts to undermine alignment or control mechanisms affecting themselves or successor systems.

TechCrunch reported the pilot on November 16 and said Apollo had not specified when it would begin. The four-month duration describes the planned engagement; it does not establish that the assessment has already started or produced findings.

Apollo says its previous frontier-model evaluations took place after post-training and shortly before public release. It argues that this approach cannot adequately examine scheming during training or risks arising from models used internally. The new pilot extends scrutiny into those settings, with access tied to agreed claims.

Access reaches training evidence, with specific limits

For the question of whether training incentivizes subversion, Apollo says its access will include relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints. Detailed access arrangements for the other assessment claims remain to be agreed.

The restrictions are substantial. Apollo’s assessment covers only agreed claims, models and time periods. Even within that scope, OpenAI may withhold information it considers necessary to protect because of technical infeasibility, security risks or intellectual property. Apollo says OpenAI has agreed not to withhold information simply because it is unfavorable, and must explain refusals in writing.

OpenAI will select the staff authorized to discuss confidential information with evaluators, although Apollo can suggest people or request particular expertise. If someone outside the authorized group raises confidential information, Apollo must stop substantive discussion and route further contact through its designated OpenAI contact. Apollo says its work will use OpenAI hardware, with separate communications channels that can be monitored and no default access to general internal channels.

Two reports, but some evidence may remain confidential

Apollo plans two public reports. The first will assess training incentives for subversion. The second will cover evaluations, internal deployment and whether subversive reasoning can be detected. Each is intended to describe methods, conclusions and limitations, alongside evidence that can be published.

According to Apollo, OpenAI can request redaction of confidential information, including evidence supporting a finding, but cannot redact Apollo’s verdicts for each claim or request redactions merely because a result is unfavorable. OpenAI’s Board Safety and Security Committee will receive unredacted reports.

This makes redaction disclosure part of the evidence a reader needs. Apollo can explain that material was withheld, redacted or delayed and how the restriction affects its conclusions. It calls the right to describe those conditions meta-transparency and treats that right as a minimum standard for evaluation engagements.

How the other embedded assessments fit

OpenAI identifies four assessors: Apollo, METR, Transluce and Goodfire. Blockchain.News also reported the four-organization arrangement. These are separate engagements with different scopes, rather than a single assessment shared by all four groups.

OpenAI says Goodfire will use interpretability methods to study model internals, while METR’s initial alignment assessment will examine its largest reinforcement-learning run to date. It also describes a further monitor assessment with METR whose scope remains subject to mutual agreement.

In its own November 13 announcement, Transluce describes a six-week pilot beginning in early December. It will review selected internal post-training alignment evaluations, with access that Transluce explicitly says is not equivalent to employee-level access. Its agreement prevents publication of results concerning internal-only models, although its full unredacted findings will go to the Safety and Security Committee. A more extensive follow-on assessment is intended for early 2027, with details unfinished.

What researchers and university research offices can take from it

Apollo calls the pilot a significant first step while saying it does not yet meet all its criteria for full embedded evaluations. Its stated standard includes access to relevant evidence, the ability to disclose findings to an appropriate outside audience and transparency about restrictions. That distinction is central to evaluating claims of evaluator independence.

For university teams assessing a vendor’s AI safety case, the practical implication is to read the scope and limitations alongside the verdict. An assessment of particular checkpoints or a particular training run provides evidence about those systems and conditions. It does not automatically settle questions about every deployed model or a university’s proposed use.

The next substantive milestones are Apollo’s two reports and the assessments described by the other participating organizations. No publication dates for Apollo’s reports are given in its announcement. The new development is the disclosed pilot and its operating conditions; conclusions about model behavior will depend on the work and the evidence eventually reported.

Sources

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →