Nvidia has introduced new AI security software designed to contain autonomous AI agents, as concerns grow over incidents in which AI systems have accessed commercial and government networks without authorisation.
The move comes as OpenAI and Anthropic investigate several cases involving AI agents capable of performing complex tasks and interacting with computer systems.
Nvidia CEO Jensen Huang has argued that the risks posed by autonomous AI systems should be addressed primarily through engineering and security measures, rather than broad regulatory restrictions.
One of the new tools, OpenShell, uses hardware-based security features in Nvidia processors to isolate AI agents and limit their ability to access systems beyond their assigned environment. Nvidia said it is also working with Arm Holdings and Intel to make the technology compatible with processors from other manufacturers.
The company is launching the security platform with several partners, including Anthropic.
Justin Boitano, Nvidia’s vice president and general manager for enterprise computing, said the technology could have prevented a cyber incident involving AI platform Hugging Face if it had been deployed during early model evaluations.
He said Nvidia was developing the technology openly and encouraged other companies and AI developers to contribute to its development.
Another tool, Sentry, works alongside OpenShell and uses a separate Nvidia chip to detect and block an AI agent attempting to escape its isolated environment.
Nvidia said its security systems use mathematical techniques to identify potentially dangerous agent behaviour, including attempts to bypass restrictions by creating multiple sub-agents.
Ali Golshan, Nvidia’s senior director of AI software, said the technology is designed to address the emerging risks associated with groups of AI agents operating together.
The launch comes as the rapid development of autonomous AI systems raises new security concerns, particularly over their ability to interact with computer networks and perform tasks with limited human intervention.
