OpenAI Partners with Four Organizations to Conduct AI Safety Evaluations
7 hour ago / Read about 0 minute
Author:小编   

On November 17, OpenAI announced a partnership with four independent organizations—Apollo Research, Goodfire, METR, and Transluce—to launch the first phase of embedded safety evaluations. Under this arrangement, external experts will be granted access to OpenAI's internal models, source code, training data, and monitoring systems so they can review model alignment and safety safeguards. In terms of division of labor, Apollo Research will focus on evaluating whether models secretly act against developer intent; Goodfire will study model internals and training monitoring; METR will examine model misalignment behavior during large-scale reinforcement learning training; and Transluce will assess the quality of safety testing during the post-training phase. Evaluators may enter OpenAI offices when necessary and access sensitive materials through dedicated devices. An independent oversight committee under the OpenAI Foundation Board will supervise the entire process and receive complete reports. The evaluating organizations retain the right to publish results independently, but OpenAI may request the removal of confidential information or delay disclosure of safety risks. OpenAI plans to further expand the scope of evaluations in 2027.