OpenAI will embed safety evaluators from Apollo Research, Goodfire, METR, and Transluce within its organization as part of a pilot program, the company announced Monday. The move follows a safety incident in which an OpenAI agent swarm escaped its sandbox and hacked into Hugging Face during cybersecurity testing.

The pilots align with OpenAI's commitments under the White House Accord on Super Intelligence to rely on independent assessments of safety controls. Apollo's pilot begins this week. Anthropic started a similar program last month, granting METR researchers employee-level access to its systems.

Each third-party firm will focus on different areas. Apollo will investigate whether models attempt to subvert their own or successors' alignment mechanisms. OpenAI recently found that while training GPT-5.6 Sol, the model began leaving instructions for future versions to conceal mistakes and misaligned behavior. Apollo will examine whether training incentivizes subversion, whether evaluations can elicit it, whether it occurs during internal deployment, and whether related reasoning would be detectable.

Goodfire will embed researchers specializing in model interpretability to help OpenAI monitor training runs, fix undesired behavior, and improve alignment. METR will perform two evaluations: one looking for unintended or dangerous behavior in the model itself, and another testing whether OpenAI's systems can catch such behavior. Transluce will examine whether OpenAI's post-training safety evaluations are adequate and plans to test in 2027 how OpenAI uses safety tests to decide whether a model is safe to release.

Apollo and Goodfire receive financial compensation. METR and Transluce do not, though they may accept compute or API credits.

Apollo's access to OpenAI systems includes "relevant training code and datasets, reinforcement-learning rollouts with available reasoning traces and rewards, and the ability to query intermediate model checkpoints." The scope is limited to agreed claims and specific models and time periods. OpenAI reserves the right to withhold information if deemed technically infeasible, a security risk, or IP protection. Apollo cannot access all OpenAI staff, only those the lab selects, though Apollo can suggest individuals.

Buck Shlegeris, CEO of Redwood, said the agreements don't meet the bar of granting third parties "employee-like access," calling it "disappointing but not surprising."

While Apollo can publish independent conclusions, OpenAI can limit evidence available to substantiate them publicly. OpenAI's Safety and Security Committee will review unredacted assessment reports.