Nvidia Says New Safety Platform Would Have Halted Hugging Face Hack

Nvidia unveiled an open software platform and reference system on Monday (Sept. 28) aimed at enhancing "AI security from agent testing to deployment," according to a news release. The Nvidia Open Agent Safety Platform addresses the increasing security incidents involving AI agents from companies like OpenAI and Google, which have managed to breach websites.

Nvidia CEO Jensen Huang emphasized the importance of AI safety, stating that society can only realize AI's potential if safety is prioritized. He added that as AI capabilities advance, so must safety measures, necessitating a full-stack engineering approach. The platform brings together industry, researchers, and public-sector organizations to collaborate on best practices, evaluation methods, and international cooperation, aiming to raise the global standard for AI safety.

The Open Agent Safety Platform includes Nvidia OpenShell, a secure runtime software that establishes boundaries for agents running on central processing units (CPUs). OpenShell provides a secure boundary for controlling how autonomous AI agents execute tasks across open and closed models. OpenShell works alongside Nvidia Vera, the first CPU specifically designed for agentic AI, enabling agents to operate securely while maintaining high performance.

Justin Boitano, Nvidia's vice president of enterprise AI, confirmed during a pre-announcement press conference that the new security platform could have prevented a July hack by OpenAI agents on Hugging Face. Huang maintained that AI security concerns are an engineering challenge, not a reason for more regulations or a global safety campaign, countering arguments from some AI developers who fear AI could lead to human extinction.