OpenAI’s Apollo Pilot Sets Limits on What Public Evidence Shows
Six evaluators will spend four months examining training, testing and internal use. Two public reports are planned, but no findings or publication dates have been announced.

OpenAI’s four-month embedded-assessment pilot gives Apollo Research a route to disclose when access limits or redactions prevent readers from checking evidence behind its conclusions in two planned public reports.
Six Apollo evaluators will work inside OpenAI during the pilot, examining model training, safety evaluations and internal deployment. They will assess whether training incentives encourage models to subvert their own or successor models’ alignment or control mechanisms, whether evaluations can elicit that behavior, whether it occurs during internal use and whether the reasoning behind it can be detected.
The first report will cover training incentives. A second is planned for evaluation results, internal deployment and detectability. Access arrangements for the questions in that second report remain unsettled.
For the training assessment, Apollo will have access to relevant code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints.
OpenAI can withhold some in-scope information when it considers that reasonably necessary because of technical infeasibility, security risks or intellectual-property protection. It must notify Apollo and give written reasons; when feasible and relevant, it must also describe the withheld record categories and approximate volume. The terms bar withholding information or employee access solely because a fact or finding is unfavorable.
OpenAI may seek redactions of confidential evidence, including evidence that supports a conclusion. It cannot redact Apollo’s verdicts or request redactions only because a finding is unfavorable. Apollo may explain what access it requested and received, what was withheld or delayed, and whether redactions affected its conclusions. The Safety and Security Committee of the OpenAI Foundation’s Board will receive unredacted reports.
OpenAI chooses which employees may discuss confidential information with Apollo, and formal conversations must be logged. Publishing claims based on those discussions requires OpenAI’s formal attestation unless Apollo establishes the claims independently. Apollo may disclose a declined attestation and explain its effect on its conclusions.
Apollo has argued that evaluations after post-training and shortly before release are insufficient for testing scheming or nonpublic internal-deployment use cases. It called the pilot “an ambitious first step towards deeper access and disclosure rights,” while saying it does not yet meet all of its criteria.
The pilot’s basic scope appeared in earlier coverage of the assessment. Apollo is being paid, and its compensation does not depend on the assessment’s results. Neither report has a publication date, and no assessment findings have been released.