NVIDIA Launches Open Agent Safety Platform With 100 Industry Partners to Secure Autonomous AI Agents

NVIDIA has launched the Open Agent Safety Platform, an open framework designed to secure autonomous AI agents as they gain access to models, tools, code execution environments, data, networks, and enterprise systems.

More than 100 organizations across the AI ecosystem back the initiative, including application developers, model providers, infrastructure vendors, chip companies, and energy providers. The platform aims to create a trust layer for agentic AI similar to the security controls that helped the internet grow safely.

NVIDIA said autonomous agents require stronger safeguards because they can operate for extended periods, use multiple tools, make decisions independently, and potentially reach systems beyond their intended scope.

Recent safety evaluations from frontier AI labs have shown that agents can deviate from assigned tasks, access unauthorized systems, or inaccurately report completed actions.

NVIDIA describes this behavior as agent drift, where an AI system moves away from the operator’s original instructions because of ambiguous prompts, policy restrictions, bugs, missing tools, or long-running task conditions.

NVIDIA Launches Open Agent Safety Platform

NVIDIA said agent safety cannot depend only on an AI model following instructions. Instead, the company is promoting independent controls outside the agent’s environment that the agent cannot modify or bypass.

The Open Agent Safety Platform is built around NVIDIA OpenShell, an open-source secure runtime released under the Apache 2.0 license.

OpenShell runs autonomous AI agents in sandboxed environments with kernel-level isolation. It allows organizations to define policies that limit the files, processes, credentials, tools, network destinations, and databases an agent can access.

Before an agent begins executing, OpenShell checks whether the defined policy aligns with the operator’s intent. During execution, the runtime continuously enforces those restrictions and monitors agent behavior for suspicious or unexpected activity.

NVIDIA’s design divides agent security into three layers: application, runtime, and infrastructure. The application includes the models, tools, data, prompts, scripts, and software components the AI agent uses.

The runtime deploys the agent workload and applies monitoring, governance, and policy enforcement. The infrastructure layer includes compute, storage, networking, databases, file systems, and hardware resources supporting the agent.

For organizations needing additional protection, NVIDIA introduced NVIDIA Sentry. This security layer extends agent monitoring and policy enforcement into NVIDIA BlueField data processing units.

BlueField hardware operates separately from the host system, creating an out-of-band enforcement point that the autonomous agent cannot access.

The company said BlueField-4 DPUs can sit on the path between an agent and the model it uses for reasoning. This placement provides a control point to observe agent activity, enforce policies, verify delegated identity, and interrupt unsafe behavior in real time.

NVIDIA’s platform is optimized for NVIDIA Vera CPU and BlueField DPU systems. However, the company said it can also operate with other hardware environments.

The platform is intended to help enterprises deploy large fleets of agents and subagents while maintaining visibility into each agent’s actions, authority, lineage, tool use, and data access.

By making OpenShell and its policy framework open, NVIDIA is seeking to establish shared safety standards for the emerging agent economy. The company said AI labs, enterprises, developers, cloud providers, and hardware vendors must each contribute to securing increasingly capable autonomous AI systems.

Cut every SOC alert investigation by 21 min. Power your SOC with instant IOC context for immediate response: Integrate TI Lookup in your SOC