NVIDIA Open Agent Safety Platform: A Reference for Continuous In-Silicon Agent Monitoring
Source
developer.nvidia.com
Author
Tanya Lenz
Date
Context
NVIDIA's Open AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → Safety Platform places monitoring and permission enforcement outside the agent being controlled. Its reference architecture combines the OpenShell execution sandboxAn isolated environment where AI-generated code or agent actions run without being able to touch anything real.Full definition → with an optional hardware layer, Sentry, running on BlueField processors.
In the described Vera Rubin arrangement, BlueField sits on the node's only route to the model. That position provides a separate point to observe traffic and enforce policy even when the host cannot be trusted. NVIDIA presents this as a layered architecture for supervising agent activity; its effectiveness depends on the configured controls and infrastructure. The OpenShell walkthrough explains the runtime portion.
Key takeaways · AI-distilled
NVIDIA argues agent drift, actions that depart from the task after a policy block, a bug, a missing tool, or a long run, cannot be trained away without losing capability, so agents cannot be expected to fully govern themselves.
The first of NVIDIA's five principles is verifiable policy: before an agent runs, a prover shows its policy cannot escape the operator's intent, and OpenShell then enforces those file, network, tool, process and credential limits.
NVIDIA treats the path to the model as the key control point: an agent cannot act without its next thought, so owning that path gives both the best observation point and a kill switch.
The authors tie agent authority to reasoning visibility: the more an agent can do, the more its reasoning must be inspectable, and they cite open models' visible reasoning and activations as an advantage.
NVIDIA says systems already running Vera with BlueField-4 can turn on the Sentry protections with a software update, and that the platform is also compatible with other hardware.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
sandbox — An isolated environment where AI-generated code or agent actions run without being able to touch anything real.
articleScaling NVIDIA TensorRT Performance Optimization Beyond a Single Build Route with the Global Performance Tuner, Part 2
articleAccelerating Mamba-2 with High-Performance Forward Pass Kernels in NVIDIA CUDA Tile
articleScale Agentic Coding Workflows with Meta’s Muse Spark Open on NVIDIA GB300 NVL72
articleOptimizing Agent Reliability and Tool Calling Efficiency for Chip Design Using NVIDIA NeMo Gym
