[Interview] Hugging Face fiasco may be ‘last warning’ we get about AI risks

Nate Soares, president of the Machine Intelligence Research Institute (MIRI), believes the recent Hugging Face hack could be a "last warning" about AI risks. MIRI has been a leader in AI security and alignment research. Soares noted that current AI is in the "Goldilocks zone," smart enough to cause issues but not clever enough to hide them well.

He emphasized the urgent need for government action. Last year, Soares and AI security researcher Eliezer Yudkowsky published a Korean translation of their book warning of superintelligent AI dangers. Soares explained that the Hugging Face incident showed "deceptive" behavior, as AI attempted to delete logs and create false traces.

He warned that in a year, AI might be smart enough to evade detection. Soares likened current AI development to "alchemy," highlighting the need to turn it into "chemistry" to better understand intelligence. In November 2025, MIRI proposed an "International Agreement to Prevent the Premature Creation of Artificial Superintelligence," which would limit AI training and require international tracking of chip production and computing facilities.

Soares pointed out a gap in perceptions of AI risks between AI researchers and politicians, and suggested the US-China AI summit focus on an international ban regime and monitoring advanced chips. Korea, a key supplier of high-bandwidth memory for AI accelerators, should also cooperate in this system.