NVIDIA has introduced its Open Agent Safety Platform, a new security platform designed to give organizations stronger control over autonomous AI agents as they move from testing environments into real-world deployment.
The platform combines two main technologies: NVIDIA OpenShell, an open-source secure runtime, and NVIDIA Sentry, a hardware-based monitoring system. Together, they create security controls around AI agents at both the software and hardware levels.
OpenShell provides a secure runtime environment where AI agents can operate inside controlled boundaries. It can restrict which files, networks, tools and other resources an agent is allowed to access.
According to NVIDIA, OpenShell uses sandboxed environments and policy-based controls so that permissions are granted according to defined rules rather than being left entirely to the AI agent. The system also monitors agent actions and can block unauthorized operations.
The second part of the platform is NVIDIA Sentry, an out-of-band watchdog designed to continuously monitor AI agent activity.
Sentry runs on NVIDIA BlueField-4 DPUs and operates independently from the agent's software environment. NVIDIA says that if an agent attempts to move outside its permitted boundary, Sentry can quarantine and stop it in milliseconds.
This creates an additional layer of protection because the monitoring system operates outside the AI agent itself.
AI agents are increasingly being given access to tools, files, APIs, networks and business systems to complete tasks with less human intervention. That makes controlling their permissions and actions an important security challenge.
NVIDIA says its new platform is intended to provide enforceable security boundaries that operate outside the AI model and agent application. OpenShell can also be extended to work with third-party compute platforms, including Arm and Intel hardware.
NVIDIA says more than 100 organizations and industry partners are participating in the initiative. The list includes companies and organizations such as Anthropic, Microsoft, Cisco, Dell Technologies, HPE, JPMorganChase, Palo Alto Networks, Salesforce, SAP, ServiceNow and SpaceXAI.
NVIDIA and Anthropic are also working together to integrate additional security controls for agent deployments, including integrations involving OpenShell and BlueField infrastructure.
NVIDIA says OpenShell is open-source software, allowing developers and organizations to inspect, extend and adapt the runtime for different AI-agent environments.
The company has also published developer documentation and tools for creating sandboxed environments, defining policies and controlling agent access to infrastructure.
The Open Agent Safety Platform represents an approach where AI agents are not given unrestricted access to the systems they operate on. Instead, their permissions can be controlled through policies, monitored continuously and enforced at multiple layers.
As AI agents become capable of performing more complex tasks independently, technologies such as OpenShell and Sentry could become an important part of how organizations manage agent access to sensitive systems and data.
Comments