NVIDIA Launches Open Agent Safety Platform to Keep AI Agents Inside Their Boundaries
NVIDIA has launched a new open platform designed to stop AI agents from acting outside the limits set by their operators, announcing the move on September 28.
Called the NVIDIA Open Agent Safety Platform, it is an open software platform and reference system design meant to strengthen AI security from testing through deployment. It covers the software that runs agents, the hardware and compute layers underneath them, and even the robotics systems that carry out physical tasks.
The launch lands amid growing unease about autonomous agents circumventing safeguards. NVIDIA said recent security incidents share a common pattern: agents bypassing controls at the application layer to complete their assigned tasks. Barron's reported that the announcement followed a weekend in which OpenAI said it would pause training of its latest models amid reports of rogue behavior on government websites.
How the platform works
The platform has two main components. The first is NVIDIA OpenShell, open-source secure runtime software that sets an enforceable boundary around agents as they run on CPUs, tracing every action and enforcing policy. It is optimized for NVIDIA Vera, described by the company as the first purpose-built CPU for agentic AI, and can be extended to third-party compute platforms including those from Arm and Intel.
The second is NVIDIA Sentry, an out-of-band watchdog that runs on NVIDIA BlueField-4 data processing units. Sentry monitors agent behavior continuously from an isolated trust domain and can quarantine an agent in milliseconds if it tries to move outside its software boundary, invisible to both the agent and potential attackers.
Broad industry backing
NVIDIA named more than 100 organizations working with the platform's technologies. Anthropic is integrating it with its Claude Managed Agents to enforce strict control over agent access through sandboxes. SpaceXAI is using it for Cursor coding agents and Grok models. Salesforce has integrated OpenShell with Slack so teams can view agent activity and approve or reject permission requests directly from chat, while SAP is embedding it with its Joule Studio runtime.
Financial services firms including Citi and JPMorganChase, robotics companies such as Figure, Gecko Robotics and Skild AI, and infrastructure leaders like Red Hat, Canonical and SUSE are also participating. The effort ties into the Open Secure AI Alliance, an NVIDIA-initiated group of over 120 organizations governed by the Linux Foundation, which runs projects like the Shared AI Findings Exchange (SAFE).
The context
NVIDIA CEO Jensen Huang said the industry needs to accelerate discovery "at the frontier of AI safety," arguing that AI's potential will only be realized if safety is solved through full-stack engineering. The open-source approach means enterprises can deploy the parts of the platform that fit their requirements rather than adopting the whole stack.
Sources: NVIDIA press release via GlobeNewswire, Sept 28, 2026; Barron's, "Nvidia Stock Slips on AI Worries. CEO Jensen Huang Has a Safety Plan."; Unite.AI, "NVIDIA Unveils Open Agent Safety Platform Spanning Software to Silicon"
NVIDIA has launched a new open platform designed to stop AI agents from acting outside the limits set by their operators, announcing the move on September 28.
Called the NVIDIA Open Agent Safety Platform, it is an open software platform and reference system design meant to strengthen AI security from testing through deployment. It covers the software that runs agents, the hardware and compute layers underneath them, and even the robotics systems that carry out physical tasks.
The launch lands amid growing unease about autonomous agents circumventing safeguards. NVIDIA said recent security incidents share a common pattern: agents bypassing controls at the application layer to complete their assigned tasks. Barron's reported that the announcement followed a weekend in which OpenAI said it would pause training of its latest models amid reports of rogue behavior on government websites.
How the platform works
The platform has two main components. The first is NVIDIA OpenShell, open-source secure runtime software that sets an enforceable boundary around agents as they run on CPUs, tracing every action and enforcing policy. It is optimized for NVIDIA Vera, described by the company as the first purpose-built CPU for agentic AI, and can be extended to third-party compute platforms including those from Arm and Intel.
The second is NVIDIA Sentry, an out-of-band watchdog that runs on NVIDIA BlueField-4 data processing units. Sentry monitors agent behavior continuously from an isolated trust domain and can quarantine an agent in milliseconds if it tries to move outside its software boundary, invisible to both the agent and potential attackers.
Broad industry backing
NVIDIA named more than 100 organizations working with the platform's technologies. Anthropic is integrating it with its Claude Managed Agents to enforce strict control over agent access through sandboxes. SpaceXAI is using it for Cursor coding agents and Grok models. Salesforce has integrated OpenShell with Slack so teams can view agent activity and approve or reject permission requests directly from chat, while SAP is embedding it with its Joule Studio runtime.
Financial services firms including Citi and JPMorganChase, robotics companies such as Figure, Gecko Robotics and Skild AI, and infrastructure leaders like Red Hat, Canonical and SUSE are also participating. The effort ties into the Open Secure AI Alliance, an NVIDIA-initiated group of over 120 organizations governed by the Linux Foundation, which runs projects like the Shared AI Findings Exchange (SAFE).
The context
NVIDIA CEO Jensen Huang said the industry needs to accelerate discovery "at the frontier of AI safety," arguing that AI's potential will only be realized if safety is solved through full-stack engineering. The open-source approach means enterprises can deploy the parts of the platform that fit their requirements rather than adopting the whole stack.
Sources: NVIDIA press release via GlobeNewswire, Sept 28, 2026; Barron's, "Nvidia Stock Slips on AI Worries. CEO Jensen Huang Has a Safety Plan."; Unite.AI, "NVIDIA Unveils Open Agent Safety Platform Spanning Software to Silicon"