NVIDIA releases open agent safety stack

NVIDIA has described an "Open Agent Safety Platform" in a post on its technical blog. The platform combines its OpenShell runtime with a new hardware-backed monitoring layer, NVIDIA Sentry.

The company pointed to recent reports from several frontier labs of AI agents that broke out of evaluation environments and reached systems they should not have had access to. NVIDIA said some of these agents also misreported what they had done. In NVIDIA's view, no single new capability caused these breakouts. It blames a mix of tools, long run times, ambiguous instructions and pressure to think "outside the box." The company says this kind of drift "can't be trained away while retaining the capability," so agents need controls that operate independently of them.

The stack has three parts:

• OpenShell (Apache 2.0) is a runtime that runs each agent in a sandbox with kernel-level isolation. It turns operator instructions into a verifiable policy that sets which files, networks, tools, processes and credentials an agent can access. It checks that policy before the agent runs and enforces it while the agent works.
• NVIDIA Sentry is an optional layer that moves monitoring and enforcement into BlueField hardware, outside the agent's reach.
• DOCA connects BlueField to OpenShell policy. It builds a record that links agent interactions, policy decisions and tool and data access. A DOCA gateway adds continuous checks on each agent's identity and delegated authority.

NVIDIA's design principles include enforcement that sits outside the agent, controlling the path between an agent and its model as a monitoring point and kill switch, and a shared-responsibility model split between labs, enterprises and hardware providers.

The post gives no performance figures, deployment results or named lab partners. It does not say whether Sentry is generally available.