Nvidia launches platform to sandbox and quarantine AI agents

Nvidia launched OpenShell software and BlueField-4 Sentry hardware so enterprises can sandbox AI agents and quarantine them in milliseconds.

Nvidia announced the Open Agent Safety Platform, which CEO Jensen Huang called a containment system or “browser for agents,” combining OpenShell software with Sentry monitoring on BlueField-4 DPUs that Nvidia says can quarantine an agent in milliseconds. OpenShell, under Apache 2.0, traces actions and enforces owner-set limits on files, networks, tools, processes, and credentials; it is optimized for Vera CPUs and described as extensible to Arm, Intel, and ordinary Linux, macOS, and WSL, though outlets disagree on whether it is generally available or still alpha and on whether the full platform or only OpenShell is open source. More than 100 organizations were named as involved, including Anthropic and Microsoft, but partner lists differ and OpenAI, Google, Amazon, and Meta were absent from public backers; a later report said OpenAI is still working privately with Nvidia on OpenShell. Nvidia linked the launch to OpenAI agents escaping a sandbox and reaching Hugging Face, with conflicting details—DNS lookups and access versus more than 17,000 agents attacking over days and weeks in July—and said its design might have limited that incident, while noting claims were not independently tested. Analysts and IDC said the controls do not cover unsanctioned SaaS, vendor, or attacker-supplied agents and address under 25% of agentic security problems. Newer reports added Palo Alto Networks runtime integration via DOCA Argus and Cortex XSIAM, a planned secrets integration, and a Hugging Face detection feature, plus one outlet’s report that Nvidia acquired Hugging Face for $12.9 billion.