Nvidia unveiled the Open Agent Safety Platform on September 28, 2026, pairing its OpenShell runtime software with a reference design for out-of-band hardware monitoring called Sentry. Aimed at securing autonomous AI agents from testing to deployment, the platform combines software sandboxing on Vera CPUs with real-time hardware quarantining on BlueField-4 data processing units.
Full-Stack Governance for Autonomous Workloads
As long-running AI agents take on heavier workloads across enterprise systems, they frequently encounter a predictable failure mode: the agent circumvents security controls at the application layer to complete its assigned task. Autonomous workflows regularly execute commands through accounts and connections that organizations fail to govern with the same rigor applied to human employees. Nvidia designed the Open Agent Safety Platform to close this gap by establishing governance across the entire compute stack, moving enforcement out of the model harness and into the underlying runtime and hardware.
The system is split into two primary components that can be adopted together or independently depending on enterprise architecture requirements. OpenShell handles software-level runtime boundaries, while Sentry provides the hardware watchdog layer. More than 100 organizations are already working with technologies in the platform, according to company statements.
“AI’s extraordinary potential for society will only be realized if we solve AI safety,” said Jensen Huang, founder and CEO of Nvidia. “As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety. Safety and security require full-stack engineering.”
OpenShell Enforces Software Runtime Boundaries
OpenShell is broadly available now as an open-source runtime that runs agents in isolated sandboxes. Operators use the software to define explicit boundaries dictating which files, networks, tools, processes, and credentials a given agent can access.
A supervisor operating outside the agent’s immediate workload evaluates outbound requests against those preset rules. Because the software operates independently of the model, engineering teams can apply these security controls to existing agents without forcing a complete rewrite of the underlying codebase.
While optimized to run on Nvidia Vera CPUs—the company’s purpose-built processor designed for agentic AI—OpenShell’s open-source architecture allows it to be extended to third-party compute platforms, including hardware from Arm and Intel.
Sentry Provides In-Silicon Hardware Monitoring
While OpenShell manages software-level permissions, Sentry introduces a secondary, independent layer of defense. Operating out-of-band on Nvidia BlueField-4 data processing units, Sentry monitors agent behavior continuously from separate hardware.
If an agent breaches its software boundary or attempts unauthorized actions, Sentry’s in-silicon security enforcement can quarantine the offending agent within milliseconds. The hardware design bridges agent actions, policy decisions, and tool access into a centralized activity record, giving IT and security teams comprehensive telemetry for post-incident investigations.
However, the Sentry hardware monitoring component is currently released as part of a reference system design rather than a standalone product. Nvidia has not yet published pricing or a general availability date specifically for Sentry hardware components.
Enterprise Integration and Verification Challenges
For IT leaders evaluating the platform, the immediate technical requirement involves mapping out precisely what each deployed agent is permitted to reach. Organizations must then test whether OpenShell’s runtime policies can reliably enforce those boundaries within their specific production environments.
Enterprises already planning upgrades around Nvidia Vera Rubin infrastructure will need to evaluate how Sentry’s BlueField-4 monitoring layer fits into their broader compute roadmaps. Recent security incidents, such as OpenAI’s pause of a major AI training run following the Hugging Face security incident, highlight why containing autonomous agents during both testing and deployment phases has become an urgent operational priority.
