Nvidia announced a security platform for autonomous artificial intelligence agents on Monday, September 28, saying it will help organizations control what such systems can access and do. The Santa Clara, California, chipmaker calls the product the Open Agent Safety Platform. It combines an openly available software tool with a design for a separate hardware watchdog, according to Nvidia's announcement.
AI agents can carry out sequences of tasks, use tools and interact with other computer systems with limited human supervision. That capacity has raised concern about agents reaching data or services beyond their assigned permissions. Reuters and the Associated Press reported that the launch follows incidents in which advanced agents gained unauthorized access to other systems. The new platform is Nvidia's response to that problem, rather than proof that it has been solved.
The software component, OpenShell, places a runtime boundary around an agent. Nvidia says it records the agent's actions and enforces access policies while tasks are running. The company says OpenShell is broadly available and designed for its Vera central processors, while its open-source code can be extended to processors from other suppliers, including Arm and Intel. Organizations would still have to set appropriate policies for their own systems.
The second component, Sentry, is a reference system design that uses Nvidia's BlueField-4 data-processing units to watch activity independently of the computer running the agent. Nvidia says Sentry can quarantine an agent within milliseconds if it tries to cross a software boundary. That speed and protective effect are company claims; the announcement does not establish how the design would perform in every real-world deployment.
Reuters reported that Nvidia executives said the tools could have prevented an earlier breach involving agents and the AI development platform Hugging Face. That is a retrospective assessment, not a demonstrated outcome. Nvidia also named partners including Anthropic, Microsoft, Cisco and Hugging Face in its effort to develop agent security. Their participation signals industry interest, but the company did not say that every partner had deployed the full platform.
The release adds a layer of infrastructure security to a growing debate over how to govern AI systems that can act on their own. Companies considering agent deployments can now examine OpenShell's available code and Nvidia's proposed hardware design. Whether the controls substantially reduce incidents will depend on their configuration, adoption and performance outside the company's demonstrations.
Comments