金色财经|Nov 16, 2026 23:18
[OpenAI Partners with Four Organizations for AI Security Assessment] According to Jinse Finance, on November 17th, OpenAI announced a collaboration with four independent organizations—Apollo Research, Goodfire, METR, and Transluce—to launch the first phase of an embedded security assessment. This collaboration allows external experts access to internal models, source code, training data, and monitoring systems to review model alignment and security measures. Apollo Research will focus on assessing whether models secretly violate the developer's intentions; Goodfire will study the model's internal mechanisms and training monitoring; METR will review model misbehavior during large-scale reinforcement learning training; and Transluce will evaluate the quality of security testing during the post-training phase. Assessors will have access to OpenAI's offices when necessary and will have access to sensitive materials via dedicated equipment. The OpenAI Foundation's independent oversight committee will supervise the entire process and receive full reports. The assessment organizations reserve the right to independently publish their results, but OpenAI may request the removal of confidential information or a delay in disclosing security risks. OpenAI also plans to further expand the scope of the assessment in 2027.
Share To
Timeline
HotFlash
APP
X
Telegram
CopyLink