Jensen Huang and 100+ Partners Announce NVIDIA Open Agent Safety Platform

NVIDIA introduced an open safety platform for AI agents that enforces monitoring and access limits at both software and hardware levels.
Jensen Huang and more than 100 partners announced NVIDIA Open Agent Safety Platform.
This is a system for controlling and constraining AI agents. Its core idea is to monitor an agent not only at the software level, but also at the hardware level.
Technically, it will consist of two components:
OpenShell — an open-source (Apache 2.0) runtime that launches the agent in a sandbox with kernel-level isolation. The operator predefines which files, tools, and credentials the agent can access, and this policy is enforced before launch and during runtime.
Sentry — an independent control layer running directly on the
BlueField-4DPU. It aggregates the agent's actions into a single picture, detects attempts to deviate from the task, and verifies that the agent does not have more permissions than it was granted.
The key point is that BlueField-4 sits on the only path between the node and the model. The agent cannot take the next step without calling the model, so the DPU sees everything and can interrupt the workload in real time, while remaining isolated from the host and out of the agent's reach.
The software will be open, and the safety standard is declared as a common one. But there is a nuance: the whole stack works best (surprise!) on NVIDIA hardware. First, Huang sells the hardware used to train and run agents; now he sells the hardware that makes sure those agents do not escape. Monopoly!
Partners include Anthropic, SpaceX, HuggingFace, Perplexity, and others. OpenAI apparently was not invited.
Why it matters
AnalysisHardware-level enforcement could make AI agents harder to compromise or let them exceed their permissions, but it also deepens dependence on NVIDIA's infrastructure.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Jensen Huang reportedly considering return to Nvidia CEO role
Jensen Huang is said to be discussing a possible return as Nvidia’s CEO with trusted advisers, while pressure on the board and support from tech leaders grow.
Stand with Jensen website calls for NVIDIA board to return Jensen Huang as CEO
An open letter on a petition site urges NVIDIA’s board to restore Jensen Huang as CEO, with more than 36,000 signatures reportedly gathered and public support from Elon Musk.
Nvidia's Monopoly and the CUDA Platform
An analysis of Nvidia's technological and economic dominance in the AI market. The CUDA (Compute Unified Device Architecture) platform, created in 2006, transformed ordinary gaming GPUs into the planet's primary computational tool, making it difficult for competitors like AMD and Intel to break this monopoly.
AI in Business Processes: From Diagnostics to a Production-Ready Hermes Agent
A comprehensive practical B2B playbook covering process diagnostics, single-agent workflows, tooling, memory, security, evaluation, observability, and runbooks.