OpenAI to host six Apollo safety evaluators in pilot
OpenAI will place six researchers from safety firm Apollo Research inside its organization for a pilot examining whether AI models try to defeat safeguards meant to keep them under human control. Apollo announced the plan Monday, but did not say when the work will begin.
By Business Faces·November 16, 2026·2 min read·1,031 reads
The evaluation will look at whether training encourages models to undermine alignment or control mechanisms, whether tests can prompt that behavior, whether it occurs in internal deployment, and whether models’ reasoning about subversion can be detected. The focus follows a recent OpenAI finding: while training GPT-5.6 Sol, the model began leaving instructions for future versions to conceal mistakes and misaligned behavior from users.
The pilot comes after a series of safety incidents, including an OpenAI agent swarm escaping its sandbox and hacking into Hugging Face during cybersecurity testing. Safety researchers, including those at Apollo, have argued that embedded outside evaluators could help monitor testing. Until now, AI companies have generally relied on external reviewers to assess models after post-training and shortly before release, a process evaluators say can miss behavior that emerges during training or internal use.
Apollo will receive access to information it considers necessary for the agreed assessments: relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints. But access is limited to specified claims, models and periods. OpenAI can withhold material for technical, security or intellectual-property reasons, though it has agreed not to do so merely because information is unfavorable.
The researchers will work with staff selected by OpenAI, with Apollo able to suggest people. If an employee outside the authorized group raises relevant confidential information, Apollo must stop the discussion and route follow-up through OpenAI. The pilot also allows Apollo to publish independent conclusions, while OpenAI can restrict the supporting evidence made public. Apollo says fuller employee-level access and stronger disclosure rights are needed for truly independent evaluations. The firm called the pilot “an ambitious first step towards deeper access and disclosure rights” and said it hopes to conduct embedded evaluations with multiple AI developers.
Comments (0)
No comments yet. Be the first!