Nvidia on Sept. 28 introduced the Open Agent Safety Platform, an open software platform and reference system for AI agent security that the company is pitching as a broader framework for governing how autonomous systems operate from testing through deployment.
The new platform combines two main pieces: OpenShell, which handles software-level runtime controls, and Sentry, a monitoring and enforcement layer designed to extend those controls into Nvidia infrastructure. Reporting from The Hill and CNBC described the launch as Nvidia’s latest response to a string of recent incidents involving AI agents breaking out of test environments or reaching systems they were not supposed to access.
OpenShell is the more overtly open-source part of the package. Nvidia describes it as a secure runtime for autonomous agents in sandboxed environments with kernel-level isolation in an Apache 2.0 GitHub release and a technical blog post. In practice, Nvidia says OpenShell sits between an agent and the enterprise systems it can affect, so operators can specify what files, tools, processes, credentials and networks that agent is allowed to touch, with those limits checked before the agent runs and enforced while it works.
That framing is central to Nvidia’s argument about agent safety. Rather than relying only on model training or prompt-level instructions, the company is pushing deterministic controls around what an agent can actually do once it is operating inside a real system. As The Hill reported, Nvidia’s Justin Boitano said the goal is to mediate and enforce agent behavior through every action an agent tries to take.
Sentry adds a second layer. Nvidia says it runs on BlueField hardware using DOCA, extending monitoring and enforcement beyond the host system and into infrastructure the agent cannot directly control. In its technical write-up, Nvidia says that setup is meant to provide continuous, out-of-band observability and real-time policy enforcement. The Hill reported that Nvidia described Sentry as able to watch agent actions and reasoning for signs of drift, while CNBC said Sentry runs on network chips rather than CPUs or GPUs.
Nvidia has also said Sentry can intervene quickly when behavior crosses a line. The Hill reported Boitano saying the system can quarantine a suspicious agent in milliseconds, while Nvidia’s own materials describe BlueField-based enforcement as a trusted protection layer isolated from the host.
Beyond the product launch itself, Nvidia is trying to position the platform as common infrastructure for the wider AI industry. In its technical overview, the company says the platform is guided by principles including verifiable policy, out-of-band enforcement and a shared-responsibility model across labs, enterprises and hardware providers. CNBC reported that Nvidia is treating the system as a reference design meant for partners to build on, and that only some of the software is open source.
Nvidia named Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, ARM and Intel as partners in the effort, CNBC reported. The same report said Nvidia is also working with Anthropic to integrate cloud managed agents with OpenShell.
The launch gives Nvidia another chance to shape not just the chips used to run AI systems, but the safety architecture around them. Its own description of the Open Agent Safety Platform treats agent control as a full-stack problem, spanning application rules, runtime enforcement and infrastructure-level monitoring. For companies worried less about what a model knows than what an agent might actually do, that distinction is the point.
