Jensen Huang’s Nvidia has launched a new set of software tools to make AI agents safer and prevent them from accessing systems or data without permission. The company says its new tools could have stopped the recent hack involving AI coding platform Hugging Face.
The move comes at a time when AI agents are becoming more capable.
Unlike regular chatbots, AI agents can perform tasks on their own, including writing and running code, accessing files, using tools and connecting to external services.
This also creates new security risks if an agent behaves in an unexpected way, the tech giant said.
One of Nvidia’s key tools is called OpenShell. It creates a secure environment, or sandbox, where AI agents can work without getting unrestricted access to a computer or company network.
The US-based tech giant says the system can control which files, networks and system functions an AI agent can access.
According to Nvidia vice president and general manager Justin Boitano, the technology could have prevented the Hugging Face attack if it had been used during early model evaluation at frontier AI labs. However, this is Nvidia’s assessment and does not mean the tool has been independently confirmed to have prevented the incident.
The company is also introducing Sentry, which works with OpenShell to detect and block an AI agent if it attempts to escape its protected environment.
According to the brand, its security system can also identify unusual behaviour, including situations where an AI agent creates multiple smaller ‘sub-agents’ to get around restrictions. This is becoming important as AI systems are given more freedom to complete complex tasks.
Nvidia is working with several technology partners on the new security platform, including Anthropic. The company is also working with Arm and Intel to make the technology compatible with processors beyond Nvidia’s own hardware.