Nvidia released a set of security tools for artificial intelligence agents on Monday. The company claims that this technology could have prevented the targeted attack on Hugging Face, an AI programming platform that was targeted by OpenAI agents.
This development comes at a time when both OpenAI and Anthropic are investigating multiple incidents involving agents capable of penetrating both commercial and governmental systems. Nvidia addresses the containment of this type of behavior as an engineering challenge, focusing on technical mechanisms to limit the activities of these systems.
One of the instruments presented is OpenShell, which uses the hardware resources of Nvidia processors to keep AI agents in controlled environments and restrict their actions. Furthermore, the company is collaborating with Arm Holdings and Intel to ensure that this technology is functional on the processors of these two manufacturers. The launch included the participation of dozens of partners, including Anthropic.
More details about the platform
Justin Boitano, Vice President and General Manager of Enterprise Computing at Nvidia, stated that the platform could have prevented the attack on Hugging Face if it had been used during model evaluations. Boitano informed Reuters that 'as far as we know, this new security platform could have prevented the intrusion if it had been used in frontier labs for model evaluation from the beginning.' He added that the company is progressing openly and wishes to involve everyone in the joint work.
Months after the incident, Nvidia formalized the acquisition of Hugging Face for $13 billion.
Another tool introduced is Sentry, which operates in conjunction with OpenShell and utilizes a dedicated Nvidia chip. This system has the capability to interrupt an agent if it attempts to break through the barriers established by the processor.
The tools employ mathematical equations to identify behavioral patterns that may signal attempts to circumvent imposed limitations, covering:
Ali Golshan, Senior Director of AI Software at Nvidia, clarified that the issue transcends the behavior of a single agent, also encompassing how multiple agents can interact with each other. Golshan emphasized: 'We are really talking about agentic behavior: fleets of agents and how they operate together.'
In summary, the new solutions integrate software and hardware capabilities to detect any attempt to evade security protections. Nvidia plans to expand this system to various types of processors and continue partnerships with industry companies in technological development.