OpenAI opened its models to four outside safety groups on Monday. It calls the plan embedded assessments. Apollo Research, METR, Transluce, and Goodfire each get deep access to internal systems.
Apollo will send six evaluators to sit inside OpenAI. They stay for months. Their job is narrow. They check if a model works against its own safeguards. METR will study OpenAI's biggest AI training run yet. Transluce will test how well OpenAI's own alignment checks work. Goodfire will use its own tools to look inside the models.
The move lands with real tension behind it. OpenAI fired three safety researchers in October. The company said they mishandled sensitive information. One of them, Tomek Korbak, gave a simpler reason. "Talking to METR was my job," he said after his firing.
METR is now one of the four groups OpenAI just invited back in, under contract.
The announcement pulled 1,798 likes and 415 replies on X. That is a big number for a safety post. Most safety posts from labs draw far less attention than a flashy new model launch.
The reach says something about where AI's fights sit today. A year ago, most arguments were about model speed and price. Today, a lab's own staff argue with its outside watchers. They argue about who checks the work, and how closely.
Apollo and Goodfire take payment for this work. METR and Transluce skip payment, though they may take compute credits instead. OpenAI says the payment stays separate from what the groups find.
That separation matters most of all. A paid referee who only earns pay for a clean result stops being a referee. OpenAI is betting that readers, and its own shaken safety staff, will trust this arrangement is different.