NVIDIA has launched an open software and hardware reference platform to enforce security boundaries around AI agents.
The NVIDIA Open Agent Safety Platform combines OpenShell, an open-source security layer for AI agents, with Sentry, a monitoring system that runs on NVIDIA BlueField-4 DPUs.
The launch follows a series of incidents involving OpenAI, Anthropic, Google, and others wherein AI agents broke out of testing environments and gained unauthorised access to real-world systems.
OpenShell creates a security boundary around agents running on Vera CPUs. It tracks what agents do and enforces rules governing their access to data, tools, APIs, and other services.
NVIDIA said OpenShell works with both open and closed AI models and can also be extended to third-party computing platforms, including those from Arm and Intel.
Sentry adds another layer of hardware-level protection. Running separately on BlueField-4 DPUs, it continuously monitors agent activity from an isolated system and can quarantine an agent if it tries to move beyond its permitted boundaries, NVIDIA said.
The system uses NVIDIA’s DOCA software to inspect agent requests and responses, verify their identity, track their activity, and enforce access controls over data, tools, APIs, and services.
“As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety,” Jensen Huang, NVIDIA's CEO, said in a statement.
“NVIDIA Open Agent Safety Platform brings together industry, researchers and public-sector organisations to share best practices, align on evaluation methods and foster international cooperation.”
Several companies across the AI ecosystem have come together to support this initiative. This includes Microsoft, Anthropic, Oracle, SpaceX, IBM, and others. Notably, OpenAI was not included in the list of companies shared by NVIDIA.
Anthropic is integrating OpenShell and BlueField with Claude Managed Agents, while SpaceXAI is using the platform with Cursor coding agents and Grok models.
Scale AI is incorporating the technology into its agentic infrastructure, while Salesforce has integrated OpenShell with Slack so users can review agent activity and approve or reject requests for additional permissions.
The platform is also being adopted by companies including Microsoft, SAP, ServiceNow, Palantir, CrowdStrike, Cisco, Palo Alto Networks, Red Hat, Hugging Face, and Perplexity. Robotics companies including Figure, Gecko Robotics, and Skild AI are using OpenShell to add security controls to systems that allow AI agents to act in the physical world.
In light of frontier models breaking out of sandboxes and leaking sensitive information, Huang has dismissed calls for new AI-specific regulation, arguing that AI safety is primarily an engineering problem that requires appropriate security frameworks.
He has also criticised “exaggerated” narratives around AI, arguing that policymakers should not give in to speculative scenarios at the expense of addressing concrete engineering and security problems.
The Open Agent Safety Platform, including OpenShell and skills, is available through the NVIDIA developer resources and GitHub.
Your reaction
Discussion
What's your take on this story?
No comments yet — be the first to share your view.