Research read

Apollo Research will embed six evaluators inside OpenAI for four months

Apollo Research and OpenAI have started a pilot in which six Apollo evaluators work inside OpenAI to test whether its models try to subvert their own oversight. Apollo says the access terms fall short of its full criteria but allow it to publish verdicts OpenAI cannot redact.

AE
Aidose Editorial
Nov 18, 2026

Apollo Research announced on November 16 a pilot in which six of its evaluators will work inside OpenAI for four months. Their task is to test whether OpenAI's models attempt to "subvert their own or their successors' alignment or control mechanisms," across training, evaluations and internal deployment.

The pilot covers four questions: whether training incentivizes subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether reasoning relevant to subversion would be detectable. For the training question, Apollo says it will get relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints. Access for the other three questions has not yet been agreed.

OpenAI can withhold items for technical infeasibility, security risk or IP protection, but must disclose that withheld items exist and give written reasons, and has agreed not to withhold information because it is unfavorable. Evaluators work on OpenAI hardware through separate monitored channels, and OpenAI selects which staff take part in formal, logged conversations.

Apollo plans to publish two reports, the first on training incentives and the second on evaluations, internal deployment and detectability. OpenAI may request redaction of confidential evidence but cannot redact Apollo's verdicts, and unredacted reports go to OpenAI's Board Safety and Security Committee. Apollo can also publicly describe what access it received, what was withheld and any pressure it believed shaped its work, without prior OpenAI review.

Apollo says the pilot "doesn't yet fulfill" its own published criteria for embedded evaluations. Even so, it is one of the more detailed public agreements on third-party access to a frontier lab's training process, and the published terms give other labs and evaluators a concrete reference point for what such access can include.

AE
Aidose Editorial
Nov 18, 2026 · read
Back to News