Apollo Research Embeds Six Evaluators Inside OpenAI for Subversion Pilot
Apollo Research will embed six evaluators inside OpenAI for four months to test whether its models subvert alignment during training, evaluations and internal deployment, it announced on November 16.

Why it matters
- Apollo Research announced on November 16, 2026 that six evaluators will work inside OpenAI for four months.
- The pilot examines four questions, including whether training incentivizes subversion of alignment or control mechanisms.
- OpenAI may withhold material for security or IP reasons but not because it is unfavorable, and cannot redact Apollo's verdicts.
- Apollo will publish two reports and can disclose withheld material, access disputes and pressure under meta-transparency rights.
- OpenAI will share unredacted reports with its board's Safety and Security Committee.
Apollo Research will place six of its evaluators inside OpenAI for four months to investigate whether the company's models attempt to subvert alignment or control mechanisms during training, evaluations and internal deployment, the AI safety organization announced on November 16, 2026.
The pilot marks a structural shift in third-party scrutiny of frontier AI developers. Apollo's previous evaluation work for OpenAI took place after post-training and shortly before public release. Apollo considers that timing insufficient for testing misaligned behavior such as scheming, and inadequate for examining non-public uses during internal deployment. Embedded evaluators with employee-like access are intended to overcome both limitations.
The stakes are concrete. As frontier models acquire longer planning horizons and broader internal use inside the companies that build them, the question of whether a model quietly works against its own oversight moves from theory to operational risk. Apollo's pilot is an early institutional test of whether external researchers can inspect that risk at the point where it forms — inside the training pipeline — rather than at the end of it.
What will the evaluators investigate?
The six-person team will examine four questions:
- Whether training incentivizes subversion of alignment or control mechanisms.
- Whether evaluations can elicit such subversion.
- Whether subversion occurs during internal deployment.
- Whether reasoning relevant to subversion would be detectable.
The scope covers attempts by a model to undermine its own alignment or controls, as well as those of its successors. Apollo describes the pilot as an ambitious first step rather than a full embedded evaluation.
What access does Apollo get?
For the training-incentives assessment, OpenAI committed to provide 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 three assessment areas remain to be agreed.
The assessment is confined to agreed claims, models and time periods. OpenAI may withhold in-scope items when it deems that withholding reasonably necessary because of technical infeasibility, security risks or intellectual-property protection. It may not withhold information or access to staff simply because the material is unfavorable. When OpenAI withholds items, it must tell Apollo that the withheld items exist, provide written reasons, and — where it considers it feasible and relevant — identify withheld records by category and approximate volume. Apollo can request broader access through a designated OpenAI contact and escalate refusals, but expansion is not guaranteed.
Staff contact is controlled as well. Apollo can request people with specific expertise and suggest individuals, but OpenAI selects the authorized participants, and formal conversations must be logged. If an OpenAI employee raises confidential information in a conversation, Apollo must stop substantive discussion and route follow-up through its designated OpenAI contact — even when the information falls within the engagement's scope.
The evaluators will work on OpenAI hardware, using separate channels that can be monitored. By default, they will not have access to general staff channels such as Slack. Apollo notes that informal conversations and Slack can surface relevant information an evaluator might not have thought to request, and it wants such information brought into scope in future assessments.
What will the public see?
Apollo plans two public reports. The first will address training incentives for subversion. The second will cover evaluations, internal deployment and the detectability of subversive reasoning. Each report will describe methods, conclusions, limitations, questions that could not be assessed, and supporting evidence that can be published.
OpenAI will review the reports and may request redactions of confidential information, including evidence supporting findings. Two limits anchor the arrangement. OpenAI may not redact Apollo's verdicts, and it may not request redactions merely because a finding is unfavorable. It can, however, redact evidence it considers sensitive intellectual property, whether or not that evidence directly contributes to a finding.
That leaves a gap for readers. The public may be unable to examine all evidence behind a conclusion, although Apollo can state whether the redactions affected its conclusions. OpenAI will share the unredacted reports with its board's Safety and Security Committee — meaning the company's own oversight body sees what the public does not.
Claims based on employee discussions carry an additional condition. They require OpenAI's formal attestation unless Apollo establishes them independently, and the requirement covers information supplied by authorized staff in formal interviews. If OpenAI declines attestation, Apollo can disclose that it was declined and explain the effect on its conclusions without revealing the underlying confidential information.
Why does Apollo insist on meta-transparency?
Separate meta-transparency rights allow Apollo to describe publicly the engagement's scope and conditions, the access it requested and received, material withheld or delayed and the reasons given, requested report removals and their significance, and any pressure it believed shaped its work. Apollo can discuss its access and publication terms without prior OpenAI review.
Apollo frames those rights as the minimum standard for any evaluation engagement it accepts. The reasoning is direct: undisclosed restrictions can distort the meaning of a clean result. Without meta-transparency, an absence of reported problems can appear reassuring even when it reflects limits on what investigators could examine or disclose.
The pilot's conclusions will therefore need to be read alongside Apollo's account of its access and publication limits. The same verdict — no subversion found — can carry different weight depending on what the evaluators were able to inspect and what they were permitted to say. Apollo's own general standard for full embedded evaluation goes further than this pilot provides: access to all relevant records even if a developer tries to hide unfavorable information, disclosure of findings to an appropriate outside audience even if the developer objects, and sufficient public explanation of constraints for readers to understand what a report can and cannot establish.
The four-month pilot, with its withheld-material disclosures, its two public reports and its board-level unredacted copies, will supply the first real-world test of whether that standard can operate inside a frontier lab — and of how much daylight remains between what evaluators conclude and what the public gets to verify.
Source: unite.ai
More from Rebecca Stone
Show full bio
Correspondent covering companies, deals and strategy in artificial intelligence and machine learning at AI In Context.
728 articles
Related articles
- OpenAI Signs Four Independent Assessors for Embedded Safety Reviews
- OpenAI and Hugging Face reveal findings from model evaluation security incident
- OpenAI Lays Out Rules for Independent AI Safety Audits
- Anthropic Grants Outside AI Watchdog METR "Employee-Like Access"
- OpenAI details how external testers probe its frontier models