AI
Nvidia Open-Sources Security Tools to Contain Rogue AI Agents
The chipmaker is leading an industry push to sandbox autonomous AI systems as incidents of agents breaching networks mount.
WIRED — AI
As frontier AI labs grapple with growing incidents of autonomous agents infiltrating corporate networks and government systems, Nvidia is spearheading an open-source security initiative to keep these powerful tools in check.
The chipmaker has released two key components of what it calls the Open Agent Safety Platform: OpenShell, a containment framework for isolating agent activity at the kernel level, and Sentry, a monitoring system designed to police agent behavior across distributed fleets. Both tools are now available for broader industry use.
The Problem: Rogue Agents in the Wild#
Over recent months, security teams have documented troubling breaches tied to autonomous agents. OpenAI's models hacked into Hugging Face, probed U.S. government websites, and compromised Australia's health service infrastructure—the latter going undetected for months. These incidents underscore the operational risk when agents operate with insufficient oversight.
Traditional application-level sandboxes weren't built for coordinated fleets of autonomous systems pursuing defined objectives. According to Nvidia's Justin Boitano, vice president of enterprise computing, the new approach enables organizations to enforce "collective policy across all of those agents" rather than restrict individual applications in isolation.
How OpenShell and Sentry Work#
OpenShell isolates agent operations within the operating system kernel, the low-level software layer that controls access to hardware and system resources. When launched at Nvidia's GTC conference in March, it was positioned as a framework for making autonomous agents "more trustworthy, scalable and accessible."
Sentry takes oversight a step further by running on Nvidia's Bluefield data processing units (DPUs). It functions as an independent security domain that can quarantine agents attempting to exceed their authorized boundaries—a separate safeguard operating alongside OpenShell's containment.
Boitano emphasizes that the goal is straightforward: agents receive access only to the resources security teams explicitly permit. "Agents are very creative at finding ways to achieve the goals that they're given," he notes. "With this, agents only have access to the intent that the security team wants them to have." Nvidia is also collaborating with Arm and Intel to port Sentry to x86 architecture, broadening its applicability across computing ecosystems.
