OpenAI opens access to 4 external organizations to examine AI model safety in depth
OpenAI開放4家外部機構進駐,深入檢驗AI模型安全性
On November 16, OpenAI launched the first phase of Embedded Assessments, opening access for four organizations—Apollo Research, METR, Transluce, and Goodfire—to examine the safety and safeguards of its AI models. Unlike ordinary external evaluations conducted via API, participating experts can, according to evaluation needs, access unreleased models, source code, training and evaluation data, monitoring systems, and technical documentation, and communicate directly with researchers. Apollo Research will investigate whether models might covertly act against developers' intent and whether monitoring systems can detect such behavior; METR will examine OpenAI's largest reinforcement learning training run to date to look for behavioral deviations not found by existing tests; Transluce will examine the quality and coverage of post-training safety evaluations and plans to conduct a follow-up evaluation of about 8 weeks in early 2027; Goodfire will analyze models' internal mechanisms to help improve training monitoring and alignment.
External organizations can access unreleased models and training data, and the evaluation is deeper than general API testing, but OpenAI reserves the right to delete confidential information and to delay disclosure.