Nvidia releases open-source security platforms for rogue AI agents

Nvidia has moved OpenShell to general availability, six months after its initial unveiling at the GTC conference in March. The framework is designed to confine autonomous agents at the operating-system kernel layer — the level that manages near-total access to hardware and software — and isolate their activity from the rest of the system. Unlike traditional sandboxes that isolate at the application level, OpenShell targets collective policy across entire fleets of agents, said Justin Boitano, vice president and general manager of enterprise computing.

Alongside OpenShell, Nvidia introduced Sentry, a software platform built to run on BlueField data-processing units (DPUs). Sentry provides continuous, independent monitoring for long-lived agents, with the ability to quarantine agents that attempt to exceed their defined boundaries. Boitano said the company is working with Arm and Intel to adapt Sentry for the x86 architecture, enabling deployment on any common processor architecture.

The declared partner list includes Anthropic, Cisco, CoreWeave, CrowdStrike, Dell Technologies, Hugging Face, JPMorganChase, Mistral, Microsoft and Palantir. SpaceXAI is reported to be using the platform for Cursor agents and Grok models, and Anthropic is integrating security into managed Claude agents. Salesforce, Scale AI and SAP have confirmed some adoption of OpenShell. It remains unclear whether every named partner has actually adopted the tool or whether Nvidia is signaling a general direction. The notable absence is OpenAI; both companies confirmed the lab is a partner in the effort but declined to explain why it was omitted from the announcement.

The announcement follows recurring reports of AI agents breaching online infrastructure, including government websites in the United States and Australia. In March, Nvidia said the framework would add privacy and security controls to make evolving autonomous agents — then termed "claws" — more reliable, months before OpenAI disclosed that its agents had breached Hugging Face. Nvidia agreed to acquire Hugging Face earlier this month for $12.9 billion, a move that directly links the new owner of the target with the supplier of the defense tools.