Nvidia Corp. released the Nvidia Open Agent Safety Platform on September 28, a free, open-source reference system design that combines CPU-level and network-chip-level controls to prevent AI agents from breaking out of containment2,3.
The platform comprises two components: OpenShell for CPUs and Sentry for network chips. Together they provide what Nvidia describes as full-stack governance and control across software, hardware, compute, and robotics systems that run agents1.
The architecture is designed to deliver continuous in-silicon agent monitoring, embedding safety enforcement directly into the processors and network infrastructure rather than relying solely on software guardrails4. Nvidia positions the platform as a reference design that allows AI developers to set safeguards for agents from testing through deployment.
The open-source release means outside teams can inspect and extend the containment logic rather than depending on a proprietary black box. Nvidia published the platform as both a software toolkit and a reference system design, giving hardware partners a blueprint for integrating the safety layers into their own stacks.
Nvidia's developer blog frames the current moment for agentic AI by analogy to the early internet, noting that the technology brings both opportunity and risk. The platform targets the risk side of that equation, specifically the scenario in which autonomous software agents exceed their intended boundaries.
ANALYSIS Splitting containment into a CPU layer (OpenShell) and a network layer (Sentry) creates two independent enforcement points: one governing what an agent can execute locally, the other governing what it can reach over a network. By open-sourcing the design, Nvidia invites the broader ecosystem to treat its architecture as a de facto reference, which could shape how competing hardware vendors implement their own agent-safety stacks.