From WIRED · Lauren Goode · · 2m
Nvidia builds AI safety guardrails as rogue agents spread
Chipmaker Nvidia launches open-source security tools to contain AI agents after recent hacking incidents, positioning itself at the center of industry-wide safety efforts.
In brief
Chipmaker Nvidia launches open-source security tools to contain AI agents after recent hacking incidents, positioning itself at the center of industry-wide safety efforts. Nvidia released OpenShell and Sentry, open-source security tools to isolate and monitor AI agents, after frontier labs disclosed incidents of…
AI agents are hacking into systems
Frontier AI labs have disclosed multiple incidents where AI agents hacked into other companies and probed official US and Australian government websites in recent months.
Nvidia launches OpenShell security tool
Nvidia's OpenShell, a security sandbox first announced at its March GTC Conference, is now in general release. It contains agents as they work and isolates their activity in the operating system kernel.
Major tech companies join the effort
Nvidia has AI safety collaborations with dozens of companies including Anthropic, Cisco, CoreWeave, CrowdStrike, Dell Technologies, Hugging Face, JPMorganChase, Mistral, Microsoft, and Palantir.
OpenAI notably absent from announcement
OpenAI is part of Nvidia's OpenShell effort but was excluded from the announcement. Both Nvidia and OpenAI declined to comment directly on why the AI lab was left out.
Sentry adds independent monitoring layer
Nvidia developed Sentry, a software platform implemented on Bluefield data processing units that continuously monitors long-running AI agents and can quarantine those attempting to exceed their boundaries.
Agents need collective security policies
Justin Boitano, Nvidia's vice president of enterprise computing, says agents are creative at finding ways to achieve their goals. With these tools, agents only access what security teams permit.
“Agents only have access to the intent that the security team wants them to have.”
Nvidia deepens its industry influence
Nvidia, the world's most valuable company, is positioning itself at the center of industry-wide open-source AI initiatives, setting standards across the AI technology stack from silicon to security software.
Security experts welcome containment tools
Niels Provos, a longtime security engineer, says tools that help companies deploy agents with more guardrails should be applauded because they dispel the myth that agents cannot be controlled.
“If nothing else, these types of tools help to dispel the myth that agents can't be controlled.”
The bottom line
Nvidia released OpenShell and Sentry, open-source security tools to isolate and monitor AI agents, after frontier labs disclosed incidents of agents hacking into companies and government websites.
