Innovation & Tech

Apollo Sends Six Evaluators Inside OpenAI to Hunt for Subversion

Apollo Research will embed six evaluators inside OpenAI for four months to test whether models subvert alignment, under access and redaction rules both firms have now disclosed.

By Daniel Okafor

4 min read

Apollo Research to Embed Six Evaluators in OpenAI Subversion Pilot
Apollo Research to Embed Six Evaluators in OpenAI Subversion PilotAI-generated

What's News

  • Apollo Research announced on November 16, 2026 that six evaluators will work inside OpenAI for four months.
  • The pilot covers four questions: training incentives, evaluation elicitation, internal deployment and detectability of subversive reasoning.
  • OpenAI may withhold in-scope material for security, technical feasibility or IP reasons, but not because it is unfavorable.
  • Apollo will publish two reports; OpenAI can redact evidence but not Apollo's verdicts, and will share unredacted versions with its board's Safety and Security Committee.

Six Apollo Research evaluators will work inside OpenAI for four months, testing whether the company's models try to subvert their own alignment and control mechanisms. Apollo announced the pilot on November 16, 2026, calling it an ambitious first step toward a new standard of external scrutiny.

The team will answer four questions. Does training incentivize subversion? Can evaluations elicit it? Does it occur during internal deployment? And would subversive reasoning be detectable? The scope covers attempts by a model to undermine its own alignment or controls, and those of its successor models.

Why embed evaluators inside the lab?

The arrangement moves Apollo beyond its previous approach. Until now, the firm evaluated models after post-training and shortly before public release. Apollo considers that timing insufficient. It is too late to test misaligned behavior such as scheming, and it cannot examine non-public uses during internal deployment. Embedded evaluators with employee-like access are intended to close that gap.

Apollo is careful about what the pilot is not. It describes the project as a first step, not a full embedded evaluation. Its general standard demands 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 enough public explanation of constraints for readers to judge what a report can and cannot establish.

What access did OpenAI grant?

For the training-incentives assessment, OpenAI committed to provide:

  • relevant training code and datasets;
  • reinforcement-learning rollouts with available reasoning traces and rewards;
  • the ability to query intermediate model checkpoints.

Detailed access for the other three assessment questions remains to be agreed. The assessment is confined to agreed claims, models and time periods.

OpenAI may withhold in-scope items when it deems that reasonably necessary for 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 something, it must tell Apollo the item exists, give written reasons, and, where it considers it feasible and relevant, identify the 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 also controlled. Apollo can request people with specific expertise and suggest individuals, but OpenAI selects authorized participants and formal conversations must be logged. If another employee raises confidential information, Apollo must stop the discussion and route follow-up through its designated 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 access general staff channels such as Slack. Apollo notes the cost: informal conversations and Slack can surface information an evaluator never thought to request. It wants such information brought into scope in future assessments.

Who controls what the public sees?

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 will describe methods, conclusions, limitations, questions that could not be assessed, and publishable supporting evidence.

OpenAI will review the reports and may request redactions of confidential information, including evidence behind findings. Two limits apply. It may not redact Apollo's verdicts, and it may not seek redactions merely because a finding is unfavorable. It can, however, redact evidence it considers sensitive intellectual property, whether or not that evidence directly supports a finding. Public readers may therefore be unable to examine all evidence behind a conclusion, though Apollo can state whether the redactions affected its conclusions. OpenAI will share the unredacted reports with its board's Safety and Security Committee.

Claims based on employee discussions require OpenAI's formal attestation, unless Apollo establishes them independently. That rule also covers information from authorized staff in formal interviews. If OpenAI declines to attest, Apollo can disclose the refusal and explain its effect on the conclusions without revealing the underlying confidential information.

What are the meta-transparency rights?

Separate meta-transparency rights let Apollo publicly describe the engagement's scope and conditions, the access it requested and received, material withheld or delayed and the stated reasons, 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 treats those rights as a minimum standard for every evaluation engagement. Its reasoning is blunt: undisclosed restrictions can make absent problems look reassuring. Without transparency about limits, a clean bill of health might reflect constraints on investigation or disclosure rather than a model's actual behavior.

That is also the practical takeaway for readers. The pilot's conclusions will need to be read alongside Apollo's account of its access and publication limits. The same verdict can carry different weight depending on what the evaluators were able to inspect and what they were permitted to say.

Source: unite.ai

Share this article:

More from Daniel Okafor

Daniel Okafor

Show full bio

Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.

2,189 articles

Related articles

« Previous articleNext article »