NVIDIA Open Agent Safety platform moves security beyond AI models

AI agents are being trusted with more real-world tasks, but the same autonomy that makes them useful can also create security gaps. NVIDIA is addressing that problem with the Open Agent Safety Platform, an open software platform and reference system design aimed at controlling agents from testing through deployment.

The key idea is to move security beyond the model and agent harness. NVIDIA says recent incidents have shown agents can bypass application-layer controls while completing assigned tasks, creating a need for enforceable boundaries that sit outside the agent itself.

Open Agent Safety Platform Adds a Runtime Boundary

The Open Agent Safety Platform brings governance across the software, compute, hardware and robotics layers involved in running agents. Organisations can deploy its components based on their requirements.

A central part is NVIDIA OpenShell, a secure runtime that creates boundaries around agents running on CPUs. It controls how autonomous agents execute tasks across open and closed models and is available as open source software that can also work with third-party compute platforms from Arm and Intel.

OpenShell is designed to run with minimal overhead on NVIDIA Vera, its purpose-built CPU for agentic AI.

NVIDIA Agent Safety Moves Enforcement Into Hardware

The platform also introduces NVIDIA Sentry, an out-of-band watchdog running on BlueField-4 DPUs. Instead of relying on the agent to follow its own restrictions, Sentry independently monitors behaviour and can quarantine and stop an agent in milliseconds if it moves outside its software boundary.

Built using NVIDIA DOCA software, Sentry can inspect requests and responses, verify agent identity, provide attested telemetry and enforce granular zero-trust policies around data, tools, APIs and services.

That creates a different security model for Secure AI agents, where controls can remain outside the system being controlled.

Enterprise AI Governance Becomes an Ecosystem Effort

NVIDIA is also positioning agent security as a broader industry effort. Anthropic, Salesforce, SAP, Scale AI and more than 100 organisations are working with the platform or its technologies.

The applications span enterprise software, coding agents, financial services, critical infrastructure and robotics. Salesforce, for example, has integrated OpenShell with Slack so teams can view activity and audit events and approve or reject requests for additional permissions.

Read More: 

Ingram Micro New Bengaluru Office expands India operations

D4NP Report 2026 reveals India’s nonprofit digital gap

Park+ and Akamai partnership targets cloud cost complexity

Gartner deepfake survey finds 41% of CISOs faced audio deepfakes