OpenAI plans third-party safety testing during model training

OpenAI says it will allow outside groups to conduct technical safety assessments while models are being trained and evaluated, extending access beyond prelaunch testing. The company is discussing participation with METR and Redwood Research and says highly sensitive work may occur inside its offices. OpenAI has not named a partner, established access terms or confirmed the publication rights Sam Altman described ten days earlier. The plan therefore expands the proposed timing of external scrutiny without yet guaranteeing evaluators the records, independence or disclosure authority needed to test it.