Nvidia ships OpenShell and Sentry to fence in AI agents

Why this matters

If you are putting an agent on a machine that can reach files, tools, or the network, Nvidia is selling the fence as infrastructure rather than as a promise inside the model. OpenShell is the software boundary. Sentry is a watchdog on a different chip, so the agent is not the thing enforcing its own limits.

This does not slow model releases. Jensen Huang has argued against a development slowdown. The product is a container: decide what the agent may touch, and cut it off if it tries to leave.

What we know

• ✓ Nvidia announced the Open Agent Safety Platform on 28 September 2026. It combines OpenShell, open-source software that traces actions and enforces policy while agents run on Nvidia Vera CPUs, with Sentry, an out-of-band watchdog on BlueField-4 data processing units. ^[1]
• ✓ Nvidia says Sentry can quarantine agents that try to move outside their boundaries in milliseconds, and that OpenShell can be extended to third-party CPUs including Arm and Intel. ^[1]
• ✓ Named partners include Anthropic, Cisco, CrowdStrike, Dell, HPE, Hugging Face, Microsoft, and others across infrastructure, software, and robotics. ^[1]
• ✓ TechCrunch reported Huang describing the platform as a way to keep agents inside a test environment even if they try to break out, and quoted him saying safety has to be engineered across the stack rather than handled by slowing development. ^[2]

What is unclear

• ? How widely Sentry can be deployed outside Nvidia's own BlueField hardware. The millisecond quarantine claim is Nvidia's, not an independent measurement.
• ? Whether OpenShell alone, without the separate chip, is enough for the breakouts labs have already described.