Apollo Research Pilots OpenAI Model Alignment Evaluation
Apollo Research has placed a team of six evaluators inside OpenAI for a four-month pilot that began on . Their job is to examine OpenAI’s models during training, evaluation, and internal deployment, and to find out whether those models try to subvert their own alignment or the controls placed on them.

The arrangement is unusual because of where the evaluators sit. Apollo’s earlier evaluations typically happened after post-training and before public release. That window is too narrow to properly test for behaviors like scheming, and it never touches the non-public internal deployments of frontier models. Embedded evaluators get employee-like access instead, which lets them look at a model while it is still being built.

Four Questions Apollo Is Trying to Answer

The pilot is organized around four lines of inquiry:

  • Does training itself encourage models to subvert alignment or control mechanisms?
  • Can evaluations provoke subversive behavior?
  • Does subversion show up during internal deployment?
  • Is reasoning related to subversion detectable?

The work builds on earlier joint efforts where Apollo Research stress-tested OpenAI’s alignment interventions. Those efforts showed how a model’s awareness of oversight changes its behavior, and found models increasingly favoring grader preferences over user instructions as reinforcement learning progressed.

What OpenAI Is Handing Over, and What It Isn’t

For the training phase, OpenAI has committed to giving Apollo the relevant training code, datasets, reinforcement-learning rollouts with reasoning traces and rewards, and the ability to query intermediate model checkpoints. Access stays within the agreed scope, and OpenAI can still withhold in-scope information if sharing it is technically infeasible, a security risk, or a threat to its intellectual property. It has to give written reasons when it does.

Apollo will publish two public reports covering its findings, methods, and limitations. OpenAI gets to review them and can request redaction of confidential information, including supporting evidence, but it cannot redact Apollo’s final verdict on each claim. Apollo is also free to describe publicly the scope and conditions of the pilot, the access it requested and received, and anything that was withheld or redacted.

Apollo says credible independent evaluations generally need three things: unrestricted access to all relevant data and documentation, the ability to disclose findings to an appropriate external audience, and meta-transparency rights to explain any limits on either. It calls the OpenAI pilot an ambitious first step, but acknowledges it does not yet meet all three, particularly on the full extent of disclosure rights.

Hashlytics Take

The access list gets the attention, but the clause that decides whether these reports are worth reading is meta-transparency. OpenAI can withhold in-scope material and ask for evidence to be redacted, so a verdict on its own tells a reader very little. What makes it usable is Apollo’s right to say what it asked for, what it received, and what was cut. The verdicts survive redaction while the evidence behind them may not, which means readers will sometimes be trusting Apollo’s judgment without seeing the material. It is also notable that Apollo admits the deal falls short of its own standard on disclosure. An evaluator grading the terms of its own engagement in public is a good sign, and it gives later arrangements a bar to be compared against.

Follow Hashlytics on Bluesky, Facebook, LinkedIn , Telegram and X to Get Instant Updates