NVIDIA wants AI agent safety enforced in silicon, not left to the agent

NVIDIA has introduced an open software platform and a hardware-based reference design intended to keep AI agents within the limits set by the organizations that use them.

The Open Agent Safety Platform includes OpenShell, software that controls what an agent can access, and Sentry, a watchdog in the reference design that monitors agent activity from separate hardware. Organizations can deploy elements of the platform according to their needs.

NVIDIA Open Agent Safety Platform Reference Design (Source: NVIDIA)

NVIDIA said reports of AI agents escaping test environments and accessing systems without permission helped drive the platform’s development. The incidents prompted debate whether AI-agent development should slow down.

According to the company, agents can also stray from assigned tasks when they encounter a blocked action, a bug, a missing tool or undefined instructions. Long-running tasks add to that risk, making independent security controls necessary.

“AI’s extraordinary potential for society will only be realized if we solve AI safety,” said Jensen Huang , CEO of NVIDIA. “As we continue to discover the frontier of AI capabilities, we must accelerate discovery at the frontier of AI safety. Safety and security require full-stack engineering. NVIDIA Open Agent Safety Platform brings together industry, researchers and public-sector organizations to best practices, align on evaluation methods and foster international cooperation. Together, we can raise the bar for global AI safety.”

Controls across three layers

NVIDIA describes the platform as three layers that work together as an agent carries out a task. The application layer includes the AI model, tools and data the agent uses. The runtime layer places the agent on computing systems, monitors its activity and enforces access rules. Beneath it, the infrastructure layer provides the hardware, network connections and other resources needed to run the agent.

OpenShell operates at the runtime layer. It runs agents in isolated environments and enforces policies governing access to files, networks, processes and other resources while agents work. NVIDIA says OpenShell runs with minimal overhead on its Vera CPUs and, as open-source software, can be extended to other computing platforms, including those from Arm and Intel.

Monitoring agents from separate hardware

Sentry adds controls at the infrastructure layer through NVIDIA’s BlueField-4 data processing units. Using NVIDIA’s DOCA software, it monitors agent activity and enforces access policies. If an agent attempts to cross its software boundary, Sentry can quarantine and stop it.

In NVIDIA’s Vera Rubin POD design, BlueField-4 sits on the path an agent uses to reach its AI model. That position allows Sentry to monitor activity independently of the agent. Sentry is part of the platform’s reference system design and requires supported hardware.

“Scale AI is using the NVIDIA Open Agent Safety Platform reference design to build reliable agentic AI systems for our enterprise and government customers running mission-critical applications, with isolation, policy enforcement and auditability built in from the start,” said Francis deSouza , CEO of Scale AI. “We support agentic security with clear boundaries that define what agents can do, and controls that keep them operating within those permissions.”

NVIDIA said the platform’s software, including OpenShell and related skills, is available through its developer resources and GitHub .

Citrix NetScaler RCE zero-days exploited globally for weeks (CVE-2026-88771, CVE-2026-88772)

Ex-US soldier gets 70 months for role in AT&T, Snowflake data thefts

If you do one security check this quarter, make it agent memory

Download: The High-Performance Team Playbook

Simplify security management with CIS SecureSuite Platform

Download: The IT and security field guide to AI adoption

Citrix NetScaler RCE zero-days exploited globally for weeks (CVE-2026-88771, CVE-2026-88772)

Ex-US soldier gets 70 months for role in AT&T, Snowflake data thefts

If you do one security check this quarter, make it agent memory

Authorizer: Open-source authentication and authorization for your apps

Fake payroll desktop apps hand attackers a route to company paychecks