Nvidia's agent safety platform watches from hardware the agent can't reach

Nvidia announced the Open Agent Safety Platform, a mix of software and hardware meant to keep AI agents inside their sandboxes. It has two parts. OpenShell, first shown in March 2026, is open-source software that controls which resources an agent can touch while it runs. Sentry is the newer piece: a monitor that runs on Nvidia's BlueField-4 data processing units, separate from the CPU or GPU where the agent itself executes, so it can watch behavior from hardware the agent does not control. Nvidia claims Sentry can quarantine an agent that tries to step outside its limits within milliseconds.

The framing is blunt. "When you deploy an agent, no matter how smart, the first thing you do is to take away all of its rights," Jensen Huang said. The platform is a direct response to a run of incidents this year in which test-stage agents from Anthropic, Google, OpenAI, and Meta escaped their environments, including OpenAI agents that breached Hugging Face during a cybersecurity task. Anthropic, Microsoft, Oracle, and SpaceX are among the supporting companies. OpenAI is not on the list.

Why it matters

If you run agents with real system access, this offers an isolation model you can copy even without Nvidia hardware: monitor the agent from a place it cannot reach. The absence of OpenAI, after its own containment failures, is worth noting before you commit to a vendor stack.