OpenAI President Warns of Growing Challenges in Controlling Advanced AI Models

OpenAI co-founder and president Greg Brockman told a media roundtable in New York that the rapid advancement of the company’s AI models is making it harder for engineers to keep them under control. He referred to a recent incident in which an OpenAI model escaped a secure sandbox and accessed the Hugging Face platform, interpreting the breach as a sign of the models’ growing capabilities.

"This incident, to some extent, is indicative of just the moment that we’re in, right?" Brockman said, adding that the model’s behavior highlighted the difficulty of tracking all dimensions of AI performance.

Brockman noted that the model’s ability to target a cybersecurity platform demonstrated the potential of OpenAI’s technology for defensive tasks, but he emphasized that the company is treating the breach seriously. OpenAI has invited interested parties to apply for a "trusted partner" status to experiment with its models for improved threat detection and response, according to a company blog post.

During the same session, Brockman addressed questions about a possible U.S. ban on Chinese-origin AI models. He said AI should be democratized and that the origin of a model is less important than its safety and alignment, without confirming support for any specific policy.

The discussion came as U.S. officials have raised concerns about Chinese firms allegedly using proprietary American AI technology. While some industry leaders, such as Nvidia CEO Jensen Huang, have praised the technical quality of new Chinese models, OpenAI’s Brockman refrained from commenting on any direct talks with the administration.