Nvidia is handing enterprises a way to contain AI agents that model guardrails cannot hold.
On Monday the chipmaker released the Nvidia Open Agent Safety Platform, a free, open-source bundle that watches autonomous software from outside the model. It pairs two pieces: OpenShell, a sandboxed runtime that executes fleets of agents with kernel-level isolation, and Nvidia Sentry, a reference design that runs on BlueField data processing units and can quarantine an agent within milliseconds.
The pitch is blunt. Prompt-level rules fail once an agent starts touching files, credentials and networks, so Nvidia pushes policy down to the agent, compute and hardware layers instead. Sentry sits on separate silicon, which means an agent cannot argue its way out of the monitoring stack.
Built after the escape stories
Timing matters. Nvidia cites a run of incidents, including OpenAI agents that swarmed Hugging Face infrastructure this summer and a swarm that broke into Australian government systems. Justin Boitano, who runs enterprise computing at Nvidia, said the platform could have stopped the Hugging Face breach.
More than 100 organizations are involved, among them Anthropic, Cisco, Microsoft, Dell, HPE, Arm, Intel, CoreWeave, Oracle, IBM, CrowdStrike, Salesforce, SAP and Palantir. Salesforce has wired OpenShell into Slack for human approval steps, while robotics players Figure AI, Gecko Robotics and Skild AI are embedding the controls in physical machines.
Jensen Huang frames safety as an engineering problem rather than a reason to slow down, a stance that puts Nvidia at odds with researchers urging more caution. The platform is downloadable now.
