OpenAI will embed six Apollo Research safety evaluators inside its organization for four months to investigate whether advanced AI models attempt to undermine their own safety controls.
The OpenAI Apollo Research pilot will examine model behavior during training, evaluations and internal deployment. The researchers will have access to selected internal systems, although their investigations and public disclosures will face restrictions.
OpenAI Apollo Research Pilot Examines Model Subversion
In its November 16 announcement, Apollo outlined four areas of investigation. These include whether training encourages subversion, evaluations trigger it, or models exhibit such behavior during internal deployment.
Researchers will also examine whether monitoring systems can detect reasoning associated with attempts to bypass safeguards.
The pilot follows previous AI safety incidents involving unexpected agent behavior. TechAmerica recently covered OpenAI’s misalignment reports, which documented sandbox escapes, prompt injection, and other unintended actions.
Apollo has also studied whether AI models behave differently when they recognize evaluation conditions. Its research suggests that some models may prioritize evaluator preferences over their original instructions.
Apollo Researchers Receive Limited Internal Access
OpenAI will provide relevant training code, datasets and reinforcement-learning records. Evaluators can also examine available reasoning traces, training rewards and intermediate model checkpoints.
However, access is limited to agreed models, research questions and time periods. OpenAI may withhold information for security, intellectual property or technical reasons.
Apollo can interview authorized employees, but OpenAI will select the participants. Researchers cannot independently question other employees about confidential matters without approval.
The arrangement falls short of Apollo’s proposed standard for fully independent evaluations, which calls for broader access to relevant records and personnel.
Public Reports Will Explain Findings and Restrictions
Apollo plans to publish two reports detailing its findings, methods and limitations.
The first will examine whether training encourages models to undermine safety controls. The second will cover evaluation behavior, internal deployment, and detecting potentially subversive reasoning.
OpenAI may request redactions to protect confidential information. However, it cannot remove Apollo’s conclusions simply because the findings are unfavorable.
TechAmerica previously examined AI monitoring systems designed to detect rogue agents, including Apollo’s work on identifying risky actions before execution.
Apollo considers the pilot an initial step toward stronger independent oversight. The organization says broader access and disclosure rights will be necessary for comprehensive evaluations.
Neither organization has announced a specific start date. The pilot’s findings will depend on the evidence researchers can access and the agreed scope of their investigation.










