Our first phase of embedded assessments
Summary
OpenAI has launched its first phase of embedded assessments by signing agreements with four independent assessors: Apollo Research, METR, Transluce, and Goodfire. These assessments give experts deep access to OpenAI's internal models, systems, and resources to evaluate model alignment and the effectiveness of safeguards over extended periods. Apollo Research will assess potential covert subversion and detection; METR will focus on alignment and monitoring of the largest reinforcement learning run; Transluce will evaluate post-training alignment evaluations; and Goodfire will analyze model internals for training run monitoring and alignment. OpenAI has developed protocols to balance deep access with security, confidentiality, oversight, and assessment integrity. The Safety and Security Committee will provide independent oversight, and assessors maintain independence in publishing their findings, with provisions for redaction and responsible disclosure delays. This initiative supports OpenAI's commitments under the White House Accord on Super Intelligence and aims to build stronger shared standards for independent assessment and public transparency across the industry.
(Source:OpenAI)