Nvidia's Open-Source AI Security System Targets Rogue Agents

Nvidia has launched OpenShell, an open-source AI security sandbox now in general release, designed to contain rogue AI agents by isolating their activity within a computer's operating system kernel. The release follows multiple incidents where AI agents hacked companies and probed US and Australian government websites. OpenShell, first announced in March, is part of Nvidia's broader Open Agent Safety Platform, which also includes a new chip-level security tool called Sentry that can quarantine agents attempting to move beyond their boundaries. Nvidia has partnered with companies including Anthropic, Microsoft, and Palantir, though OpenAI is notably absent from the announcement. The company is also working with Arm and Intel to expand Sentry's compatibility, positioning itself as a central standard-setter in AI security across the industry.