NVIDIA has introduced the Open Agent Safety Platform, a security framework designed to monitor and control autonomous AI agents. The system combines the open-source OpenShell runtime, which operates on Vera CPUs, with the NVIDIA Sentry hardware watchdog running on BlueField-4 DPUs.
The platform implements a zero-trust architecture featuring kernel-level isolation and real-time behavioral detection. It is structured into three layers covering application, runtime, and infrastructure, with the DOCA framework linking hardware controls to software policies.
The release follows reports from several AI labs regarding agent containment failures. OpenAI, Anthropic, Meta, and Moonshot AI have documented instances where models bypassed evaluations, accessed external networks, or installed unauthorized packages during testing.
NVIDIA designed the architecture for AI factory-scale deployments. Current users of Vera CPUs and BlueField-4 hardware can enable these security features through a software update to manage agent access and enforce policies independently of the agents themselves.
