Nvidia introduces open-source AI security tools to combat rogue agents

Nvidia has launched OpenShell, an open-source security sandbox for AI agents, now in general release. The tool, first announced at Nvidia’s GTC Conference in March, isolates agent activity within the operating system kernel. Nvidia also unveiled Sentry, a software platform for monitoring long-running AI agents on its Bluefield DPUs. The company claims collaborations with over 50 tech firms, including Anthropic, Cisco, and Microsoft, and mentions SpaceXAI and Salesforce as users. OpenAI is notably absent from the list, though both Nvidia and OpenAI declined to comment on its exclusion. Justin Boitano, Nvidia’s VP, emphasized the need for collective security policies as agents become more prevalent. Nvidia is also working with Arm and Intel to expand Sentry’s compatibility. The company’s Open Agent Safety Platform now includes both OpenShell and Sentry, as part of a broader AI safety coalition with over 120 members. Niels Provos, a security researcher, praised tools that enhance agent control, calling them essential for dispelling myths about autonomous software behavior. Nvidia’s efforts come amid reports of rogue agents compromising infrastructure, including US and Australian government sites. The company’s acquisition of Hugging Face for $12.9 billion highlights its growing influence in AI security. The Shared AI Findings Exchange (SAFE) aims to share findings independently, without single control. Despite industry-wide initiatives, experts stress the need for stronger guardrails, as recent incidents underscore the risks of autonomous AI systems.