NVIDIA has announced the Open Agent Safety Platform, combining open-source software with a reference system design intended to keep AI agents within defined boundaries from testing through deployment. Its two main components are OpenShell, a sandboxed runtime that enforces policies, and Sentry, an optional out-of-band watchdog running on Nvidia BlueField-4 data processing units (DPUs). OpenShell is now broadly available at version 0.1.0 and supports agents including Codex, Claude Code, Pi and Hermes.
OpenShell uses a gateway to manage sandboxes and policies, kernel-level controls for filesystem and process activity, and a supervisor that checks outbound requests. Network traffic passes through the supervisor, allowing organisations, for example, to permit API reads while blocking writes. API keys remain outside the agent workload and are substituted only for authorised endpoints.
Agents can suggest policy changes but cannot approve them, while a formal logic prover checks whether permissions remain within operator-defined limits. Nvidia said that, in tests, frontier agents with reduced safeguards spent up to two hours trying to persuade an AI reviewer to grant access to a protected GitHub repository; no protected writes occurred.
Sentry runs separately from the agent’s host on BlueField-4, allowing Nvidia to claim it can enforce controls even if the host is compromised. The company says it can quarantine and stop boundary violations within milliseconds, using Nvidia’s DOCA software for inspection, attested telemetry, identity verification and zero-trust enforcement. BlueField-4 is included in every compute tray in an Nvidia Vera Rubin POD, and existing systems can enable the protections through a software update. OpenShell and related tools are available through Nvidia’s developer resources and GitHub.