OpenAI will allow external institutions to evaluate its artificial intelligence models at an earlier stage.

OpenAI plans to have third-party institutions review the security risks of its AI models at the early stages of their development cycle, as part of its ongoing efforts to address growing public concerns about the technology’s potential harms. The company announced in a blog post on Tuesday that it will allow external organizations to conduct technical security assessments on its models during the training, evaluation, and launch phases of new AI models. OpenAI also outlined priorities to ensure such reviews operate effectively, including "robust independence mechanisms, scientific rigor, sound security measures, and clear division of responsibilities." Lama Ahmad, who leads the company’s collaboration with external experts on security reviews, noted that previously OpenAI typically invited such groups to conduct security assessments before model release and evaluated their capabilities. "As risks rise, we need to focus not only on deployment issues but also on high-risk stages like training and evaluation," she said.