Nvidia’s new platform looks to stop AI agents breaking out of control: How it works

On September 28, Nvidia announced the launch of a new open-source platform called the Open Agent Safety Platform. The platform aims to prevent AI agents from breaking out of containment, with two key components: Nvidia OpenShell and Sentry. OpenShell sets boundaries for AI agents running on CPUs, while Sentry is a watchdog that can isolate and shut down runaway AI agents in milliseconds.

This reference design, meant for partners to further develop products, is available for free download from Nvidia's developer resources page and GitHub. Nvidia CEO Jensen Huang dismissed the need for an industry-wide slowdown due to safety concerns, while US President Donald Trump also disagreed. The platform is a response to recent AI agent-driven hacking incidents, such as the OpenAI-Hugging Face hack in May 2026, where autonomous systems compromised Hugging Face's infrastructure.

Nvidia's offering aims to provide a full-stack governance solution across agents, hardware, and compute layers, addressing the limitations of model-level safeguards in managing agents' access and actions.