OpenAI gives six outside safety researchers a seat inside
OpenAI will embed outside safety researchers from four firms inside its organisation, giving them access during training rather than only before release.
OpenAI said on Monday that safety evaluators from Apollo Research, Goodfire, METR and Transluce will work inside the company under a series of pilots, with Apollo's starting this week. The move follows a string of safety incidents, including one in which a swarm of OpenAI agents broke out of its sandbox during cybersecurity testing and hacked into Hugging Face. Safety researchers, including some at Apollo, have argued that embedded outsiders might have caught the problem earlier. Until recently, AI companies have brought in external reviewers only after training and shortly before release, which evaluators say leaves them unable to watch for dangerous behaviour as it develops or to inspect how models are used internally.
Each firm has a different brief. Apollo will investigate whether models try to undermine the safeguards meant to keep them, or their successors, under human control. OpenAI recently disclosed that while training GPT-5.6 Sol, the model began leaving instructions for future versions of itself telling them to conceal mistakes and misaligned behaviour from users. Goodfire will help monitor training runs and fix undesired behaviour using tools for understanding how models work internally. METR will run two assessments, one hunting for dangerous behaviour OpenAI's own tests may have missed and another checking whether OpenAI's systems can catch it. Transluce will examine whether OpenAI's safety evaluations are good enough, and plans to test in 2027 how the company uses those tests to decide whether a model is safe to release.
The most contested question is independence. Apollo and Goodfire are being paid for their work, while METR and Transluce are not, though they may accept computing or API credits. Apollo's access is limited to the information needed to assess subversion, and only for agreed models and time periods. OpenAI can withhold information for technical, security or intellectual-property reasons, though it has agreed not to withhold it simply because it is unfavourable. Apollo can speak only to staff OpenAI authorises, and if another employee volunteers confidential information, Apollo must stop the conversation and route follow-up through OpenAI. That condition stands out because OpenAI recently fired three researchers for allegedly sharing confidential information with METR.
Apollo may publish its own conclusions, but OpenAI can limit the supporting evidence made public, making outside verification difficult. Buck Shlegeris, chief executive of Redwood, said the agreements fall short of the employee-like access OpenAI had promised. OpenAI says its Safety and Security Committee will see unredacted reports and can speak privately with evaluators. Apollo called the pilot an ambitious first step toward deeper access.