NVIDIA Launches Open Agent Safety Platform to Stop Rogue AI Agents: NVIDIA’s New Platform Enforces AI Agent Safety at the Hardware Level
NVIDIA has launched the Open Agent Safety Platform, a full-stack software and hardware reference system designed to prevent AI agents from going rogue — from testing all the way through live deployment. CNBC
The platform introduces two core components: OpenShell, an open-source secure runtime that sets execution boundaries for agents on CPUs, and Sentry, an out-of-band watchdog running on NVIDIA BlueField-4 DPUs that can detect and quarantine misbehaving agents in milliseconds.
Unlike application-layer guardrails that sophisticated agents can bypass, NVIDIA’s approach enforces policy at the silicon level; applying zero-trust access controls, attested agent identity verification, and real-time threat detection across compute and robotics systems.
Over 100 organisations back the platform at launch, including Anthropic, Microsoft, Salesforce, CrowdStrike, Palo Alto Networks, JPMorganChase, and Figure. The open-source software is available via NVIDIA’s developer resources and GitHub.
“Safety and security require full-stack engineering,” said CEO Jensen Huang. As autonomous AI agents proliferate across enterprise and industrial environments, NVIDIA is betting that the safest agents are the ones governed in silicon, not just in software.
Read our previous articles on AI and NVIDIA here.
