NVIDIA Launches Open Agent Safety Platform: OpenShell Sandboxes Agents on Vera CPUs While Sentry on BlueField-4 Quarantines Them in Milliseconds

NVIDIA has unveiled an open reference architecture aimed at making autonomous AI agents safer to deploy in production. Rather than relying on the agent's own internal guardrails, the platform enforces safety controls from the outside, treating agents as untrusted workloads that must be monitored and contained.

The design centers on two components. OpenShell is an Apache 2.0-licensed runtime that sandboxes agents according to declarative YAML policies, limiting what actions an agent can take and what resources it can touch. Sentry runs as an out-of-band watchdog on NVIDIA BlueField-4 data processing units, giving it a vantage point independent of the host system so it can observe agent behavior and quarantine a misbehaving agent within milliseconds.

This matters because agent deployments are accelerating, and organizations need practical isolation mechanisms before trusting agents with sensitive tools or data access. By moving enforcement off the agent itself and onto dedicated infrastructure hardware, NVIDIA's approach reduces the risk that a compromised or hallucinating agent can bypass its own safety layer.

The open-source licensing of OpenShell and the reference-design approach suggest NVIDIA wants to seed a broader ecosystem around agent safety standards. If adopted, it could push the industry toward infrastructure-level agent containment as a default practice, though adoption will depend on how broadly the design is embraced beyond NVIDIA's own hardware stack.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.