Details
- NVIDIA launched the Open Agent Safety Platform on September 28, 2026, combining open-source OpenShell software with the Sentry reference system design for governing AI agents from testing through deployment.
- OpenShell creates a secure runtime boundary that traces agent actions and enforces policies on NVIDIA Vera CPUs; its open-source design can also support Arm and Intel compute platforms.
- Sentry runs on NVIDIA BlueField-4 DPUs as an out-of-band watchdog, using NVIDIA DOCA to monitor activity, verify identity, protect data and enforce zero-trust policies. NVIDIA says it can quarantine agents outside their boundaries within milliseconds.
- Anthropic, Salesforce, SAP, Scale AI, SpaceXAI, Perplexity, Microsoft, Red Hat, financial institutions and robotics companies are among more than 100 organizations working with the platform or its technologies.
- The architecture extends agent controls beyond application-layer safeguards: OpenShell governs execution in software, while Sentry independently observes and enforces policies in hardware. NVIDIA’s platform builds on the Open Secure AI Alliance, which joined the Linux Foundation in September 2026.
Impact
The launch reflects a shift toward defense-in-depth controls for autonomous agents, placing enforcement outside models and agent harnesses where compromised or misbehaving systems may bypass application safeguards. Over the next 12–24 months, open runtime standards, hardware-isolated monitoring and auditability could become important procurement requirements for enterprise, government, financial-services and robotics deployments, while increasing demand for compatible infrastructure and security tooling.
