Apollo will embed six researchers at OpenAI to investigate AI subversion
The four-month pilot reaches inside model development, but Apollo says its access and disclosure rights still fall short of its standard for full embedded evaluations.
Loading page…
The four-month pilot reaches inside model development, but Apollo says its access and disclosure rights still fall short of its standard for full embedded evaluations.
Listen to this story
The four-month pilot gives Apollo access to OpenAI training code, datasets, reinforcement-learning records and intermediate checkpoints, while details for its other assessment areas remain unsettled. Its six evaluators will test whether models develop incentives or opportunities to subvert controls, including during internal use, but OpenAI can withhold material for technical, security or intellectual-property reasons. Apollo can publish its conclusions and explain access constraints, though supporting evidence may be redacted; the arrangement is a limited step toward broader embedded reviews.
Apollo plans two public reports: one on training incentives for subversion, and another on tests, internal deployment and detectability of subversive reasoning.
OpenAI’s Safety and Security Committee will receive unredacted reports, even though public versions may omit confidential supporting evidence.
Apollo must route follow-up on confidential employee disclosures through its OpenAI contact, and claims from discussions require formal confirmation unless independently established.
Apollo Research will put six evaluators inside OpenAI for four months to investigate whether AI models try to undermine safeguards meant to keep them under human control. Announced November 16, 2026, the pilot extends scrutiny across training, testing and internal use—but Apollo says the agreement still falls short of the access and disclosure rights it wants.
Apollo’s previous frontier-model evaluations happened after post-training—the later training stages that shape a model’s behavior—and shortly before public release. Apollo argues that this timing is insufficient to investigate scheming or risks from models used internally.
The pilot is part of OpenAI’s first phase of embedded assessments, with agreements covering Apollo, METR, Transluce and Goodfire. Apollo will examine four questions:
For the training assessment, OpenAI has committed access to relevant training code and datasets, reinforcement-learning records with available reasoning traces and rewards, and intermediate model checkpoints—saved versions from during training. Detailed access arrangements for the other assessment areas remain to be agreed.
Access covers only agreed claims, models and time periods. OpenAI may withhold material for technical infeasibility, security risks or intellectual-property protection, but must identify that material exists and provide written reasons. It has agreed not to withhold information or staff access simply because it is unfavorable.
OpenAI selects authorized staff for discussions, although Apollo can suggest individuals. If another employee raises confidential information, evaluators must stop substantive discussion and route follow-up through their OpenAI contact—even when the subject falls within the assessment.
OpenAI discloses that Apollo receives payment for its work, although compensation does not depend on the assessment’s results.
Apollo plans two public reports: one on training incentives for subversion, and another on evaluations, internal deployment and detectability of subversive reasoning. OpenAI may not seek redactions merely because findings are unfavorable. Its board’s Safety and Security Committee will receive unredacted reports.
Publishing claims drawn from employee discussions requires OpenAI’s formal confirmation unless Apollo establishes them independently. Apollo may disclose that confirmation was declined and explain the effect on its conclusions.
Apollo’s minimum requirement is the right to explain constraints publicly, which it calls “meta-transparency.” It can describe withheld material, redactions and any pressure it believes shaped its work, and discuss access and publication terms without OpenAI’s prior review.
Apollo’s stated next step is fuller embedded evaluations with multiple developers, with deeper access and stronger disclosure rights. This pilot is a step toward that arrangement, not its completion.
Loading discussion...
Join the conversation
Explain what would make that trust justified—or misplaced.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.