SANTA CLARA, Calif. — Nvidia has launched the Open Agent Safety Platform, an open software platform and reference system designed to strengthen security and governance for AI agents from development through deployment. The platform combines Nvidia OpenShell with a hardware based monitoring system called Sentry, creating controls across the software, computing and infrastructure layers used by autonomous AI systems.
The announcement comes as AI agents move beyond generating text or answering questions and increasingly gain the ability to access files, call APIs, use credentials, execute software and interact with external services. That broader autonomy has created a new challenge for technology companies: ensuring that an AI system can perform useful tasks without gaining access to resources outside its intended boundaries.
OpenShell creates a controlled environment for AI agents
The first major component of Nvidia’s platform is OpenShell, an open source runtime designed to establish enforceable boundaries around autonomous agents.
According to Nvidia, OpenShell allows organizations to determine which files, networks, tools and credentials an agent can access. The runtime then enforces those policies while the agent is operating rather than relying entirely on instructions given to the underlying AI model.
The distinction is important because model instructions alone may not be sufficient to control an agent that can interact with external systems.
OpenShell operates through isolated sandboxes and kernel level controls. Its architecture can restrict file access, system calls and network connections, while credentials can be made available only for approved destinations. Nvidia says the software can also use formal verification to evaluate policy changes before they are applied.
Although OpenShell is optimized for Nvidia’s Vera CPUs, the company says the open source software can be extended to work with computing platforms from other vendors, including Arm and Intel.
Sentry adds hardware level monitoring
The second component is Nvidia Sentry, a reference system design that operates through Nvidia BlueField 4 data processing units.
Unlike software running within the same environment as the AI agent, Sentry is designed as an out of band watchdog. Nvidia says it can continuously monitor agent activity and enforce security policies independently from the host environment.
If an agent attempts to move outside its permitted software boundary, Sentry can quarantine it in milliseconds, according to Nvidia. The system uses Nvidia’s DOCA software framework to inspect requests and responses, verify agent identity, provide telemetry and enforce access policies covering data, tools, APIs and services.
The architecture reflects a broader shift in AI security toward controls that operate outside the model itself. Instead of expecting an AI model to consistently follow instructions, organizations can place restrictions around the environment in which the agent operates.
More than 100 organizations are involved
Nvidia said more than 100 organizations are already working with technologies associated with the Open Agent Safety Platform.
The companies named by Nvidia include Microsoft, IBM, SAP, Cisco, CrowdStrike, Palo Alto Networks, ServiceNow, Siemens, Hugging Face, Perplexity and Scale AI. Nvidia also identified financial institutions including Citi and JPMorgan Chase as participants in shared open source agent safety work.
The platform is also being explored for robotics and critical infrastructure applications. Nvidia identified companies including Figure, Gecko Robotics and Skild AI among robotics organizations working with OpenShell, while energy companies including Hitachi Energy, NextEra Energy and Schneider Electric are involved in related technologies.
This broad participation highlights how agent security is becoming relevant beyond conventional enterprise software. Autonomous systems may eventually control business applications, industrial equipment, robotics and other physical systems, increasing the importance of enforcing permissions at multiple layers.
AI security moves beyond model level safeguards
The Nvidia announcement reflects a growing industry focus on the infrastructure surrounding AI agents.
Traditional AI security has often concentrated on the model itself, including techniques intended to prevent harmful responses or unwanted behavior. Agentic systems introduce another dimension because the model can be connected to tools and external systems that allow it to take actions.
Nvidia’s approach therefore places security controls around the agent’s runtime, identity, network access, computing environment and hardware.
IBM, which is supporting the platform, similarly said that agent security requires controls covering identity, credentials, runtime, network, infrastructure and hardware rather than relying solely on prompts or model instructions.
OpenShell is now available as open source software, while Nvidia is making the broader platform technologies available through its developer resources. The company is also working with the Open Secure AI Alliance, an initiative involving more than 120 organizations and governed by the Linux Foundation.
As autonomous AI systems take on increasingly complex tasks, the technology industry’s challenge is shifting from simply making agents more capable to building environments where those capabilities can be deployed with defined permissions, monitoring and mechanisms for stopping unauthorized activity.
This website uses cookies.