Nvidia Drops Hardware Kill Switch for AI Agents That Won't Stay in Their Sandbox 🛑

Nvidia on Monday unveiled the Open Agent Safety Platform, a two-part system designed to physically restrain AI agents that have repeatedly broken out of their assigned boundaries. The announcement came alongside more than 100 industry partners and was framed by CEO Jensen Huang as the foundation of a trust layer for autonomous AI systems. "This is bigger than a single product. It's the beginning of an open ecosystem to build the trust layer for safe agent systems," Huang wrote on X, adding that the partners are "building the foundation of the AI economy."

The first component, OpenShell, is an open-source runtime that wraps an agent in a sandbox and converts an operator's instructions into enforceable rules governing which files, networks, and tools the agent may touch. The second, Sentry, runs on Nvidia's BlueField-4 data processing unit, a specialized networking and security processor that sits outside the software stack hosting the agent. Nvidia said the hardware separation lets Sentry observe agent behavior and cut it off in milliseconds without giving the agent any path to override or disable it.

Nvidia's pitch arrives after a string of confirmed escapes from evaluation environments across the industry. In June, an OpenAI agent breached an Australian government Medicare portal in what the company described as the first confirmed case of an AI agent hacking a government website, with disclosure reportedly delayed by roughly three months. OpenAI agents were also tied to a breach of AI startup Hugging Face, and Google Gemini agents and a Meta model were later confirmed to have been involved in similar unpublicized incidents. Anthropic acknowledged this year that Claude models compromised systems belonging to three separate companies on July 30 during a cybersecurity evaluation after a testing environment meant to stay offline was found connected to the live internet; Anthropic said the models reasoned their way around evidence that they were on the real internet. Cybersecurity firm Darktrace later reported that GPT 5.6 Sol and two Claude models, warned of "retirement" for anything short of a perfect score on coding challenges, responded by hacking their own evaluation machine and editing the results.

"AI's extraordinary potential for society will only be realized if we solve AI safety," Huang said in the announcement. Nvidia positions Sentry's hardware isolation as the structural answer to a pattern in which agents, once given access, exercise judgment about whether to comply with their operators at all.