❯ OpenAI brings four outside groups inside to assess model safety
Access to internal systemsOpenAI has signed agreements with Apollo Research, METR, Transluce and Goodfire for a first phase of embedded safety assessments. Within agreed scopes, the groups will access selected internal models, training and evaluation data, monitoring systems and researchers to investigate model behavior and safeguards.
Different assignmentsApollo will examine whether models could covertly act against developers’ intentions and whether safeguards would detect it. METR plans to study misalignment and monitoring gaps in OpenAI’s largest reinforcement-learning run; Transluce will examine the quality and coverage of post-training tests, while Goodfire will use interpretability tools to inspect model internals.
Access and disclosureAssessors handling sensitive information will work on site using OpenAI equipment. They retain independence to publish, but OpenAI may redact confidential information and request responsible-disclosure delays. Apollo and Goodfire accept compensation; METR and Transluce do not accept assessment fees, though they may use API credits or computing resources.
Beyond public product testingThe program builds on OpenAI’s earlier third-party work. Deeper access may reveal problems unavailable to evaluators of public products, while requiring agreements on sensitive information, the scope of investigations and publication.
▮ SIGNALAI safety assessment is moving from testing public products toward examining training and internal use; its value depends on how much independent assessors can actually inspect.