OpenAI announces embedded evaluators to investigate AI model subversion
🎧 Listen to this story
OpenAI announced a pilot to embed safety evaluators from Apollo Research, Goodfire, METR and Transluce within the organization. The company said the initiative supports its commitment to use independent safety assessments; the projects will begin at different times, with Apollo’s starting this week. The pilot with Apollo will investigate whether training, evaluations or internal deployment encourage models to subvert alignment or control mechanisms—including those of future versions—and whether such reasoning can be detected.
Goodfire will study how the models work internally and help monitor training runs and correct unwanted behavior. METR will evaluate dangerous behavior and whether OpenAI’s systems can detect it, while Transluce will examine post-training safety evaluations and plans to test in 2027 how they inform release decisions. TechCrunch notes that limits on access and disclosure could affect independence: Apollo and Goodfire receive financial compensation, while METR and Transluce may accept compute or API credits.
Why it matters · editorial interpretation
Embedding external evaluators within the organization could connect safety testing more directly to model training, deployment and release decisions. At the same time, as TechCrunch notes, financial compensation and limits on access or disclosure could affect how independent those assessments are.
Sources
SecurityResearch