Nvidia has unveiled a suite of security tools for AI agents. The company says the software would help protect Hugging Face, which Nvidia acquired for $13 billion a few months after the platform was attacked by unmanaged agents from OpenAI.
Nvidia has unveiled a suite of security tools for AI agents. The company says the software would help protect Hugging Face, which Nvidia acquired for $13 billion a few months after the platform was attacked by unmanaged agents from OpenAI.
The release comes amid ongoing investigations by leading US labs OpenAI and Anthropic, which are investigating numerous cases where their AI agents — systems capable of performing complex, multi-step tasks — have gone out of control and gained unauthorized access to commercial and government systems.
Nvidia CEO Jensen Huang has rejected calls for strict regulatory restrictions on AI, instead suggesting that the problem of autonomous agent escape be treated as a purely engineering challenge, similar to improving the safety of cars.
The presented protection complex includes several levels of security developed in collaboration with partners (in particular, Anthropic).
OpenShell: A tool that leverages the hardware capabilities of CPUs to create an isolated environment for AI agents. Nvidia is already working with Arm Holdings and Intel to ensure compatibility with their processor architectures.
Sentry: A support system running on a separate Nvidia chip in conjunction with OpenShell. If an AI agent tries to go beyond its allocated container on the CPU, Sentry instantly blocks it from running.
According to Ali Golshan, senior director of AI software at Nvidia, the new software uses mathematical algorithms to detect circumvention maneuvers. For example, the system recognizes cases where the main AI agent tries to create a network of “sub-agents” to bypass detected restrictions and block operator control.
Nvidia’s vice president of enterprise computing, Justin Boitano, emphasized during the briefing: “Based on what we know, this new security platform could have stopped the [Hugging Face] hack if it had been used in advanced labs from the start during model evaluation. We are developing this project openly and are committed to involving everyone in the collaboration.”
