OpenAI embeds six Apollo evaluators to monitor AI subversion risks

What happened
OpenAI has agreed to host six evaluators from Apollo Research for a four-month period starting November 16, 2026. These evaluators will work inside OpenAI to examine whether its AI models attempt to bypass or undermine alignment and control measures during training, evaluation, and internal deployment. The evaluators will have employee-like access to certain internal data and staff, subject to specific access rules and limitations. This arrangement is the first time such embedded external evaluation has been conducted at a major AI company.
Key details
- Six Apollo Research evaluators will be embedded at OpenAI for four months.
- The evaluators will investigate four areas: training incentives for subversion, evaluation methods, internal deployment behavior, and detectability of subversive reasoning.
- OpenAI will provide access to training code, datasets, reinforcement-learning rollouts, and intermediate model checkpoints for the training-incentives assessment.
- Apollo will publish two public reports, with OpenAI able to request redactions for confidential or sensitive intellectual property.
- OpenAI can withhold certain information for technical, security, or intellectual property reasons, but must disclose when and why items are withheld.
Why it matters
This pilot affects OpenAI, Apollo Research, and the broader AI safety community by introducing a new level of transparency and oversight during the development of AI models. For OpenAI, it means allowing external scrutiny of internal processes and model behaviors before public release, which could impact its operational practices and reputation. For AI developers and regulators, the results and public reports may influence future standards for safety evaluation and disclosure in the industry.
Also covered by 2 sources
- OpenAI inks agreements with four independent evaluators to boost company's safety work· 2 sources
- Building trustworthy AI by fencing it in, not predicting it· MIT Technology Review
- Our first phase of embedded assessments· Open AI News
- The Biggest AI Agent Risk Is The Workaround Nobody Approved· Forbes
- OpenAI agents keep disabling monitoring; is this being handled appropriately?· LessWrong


