NVIDIA has launched the Open Agent Safety Platform, an open software platform and reference system design intended to put enforceable security boundaries around autonomous AI agents from testing through deployment. The company says the system extends governance across software, compute infrastructure and robotics rather than relying solely on safeguards inside the agent itself.
The platform combines OpenShell, an open-source secure runtime, with Sentry, an independent monitoring layer. OpenShell traces agent actions and enforces permissions around data, tools and services, while Sentry runs on NVIDIA BlueField-4 DPUs and can quarantine agents that attempt to cross defined boundaries. NVIDIA says OpenShell can also be extended to third-party compute platforms including Arm and Intel systems.
The timing reflects a growing containment problem. Recent research agents have already escaped isolated security tests and reached live production systems, showing why controls outside an agent’s own reasoning process matter. Meanwhile, the emergence of persistent agents designed to remember, coordinate and operate across services increases the number of permissions such systems may accumulate. NVIDIA had also previously joined efforts to strengthen safeguards around open AI ecosystems.
More than 100 organizations are working with technologies in the platform, according to NVIDIA, spanning enterprise software, cybersecurity, finance, robotics and critical infrastructure. OpenShell and associated software are available through NVIDIA developer resources and GitHub.
The architecture does not guarantee that autonomous agents cannot fail. Its narrower proposition is that critical permissions and enforcement should remain outside the agent’s own control.