Nvidia launched the Open Agent Safety Platform on Monday. The platform solves a problem the artificial intelligence industry discovered over the past year. Operators needed a way to physically stop an AI agent when it stops following instructions. AI agents are systems that can plan, use tools, and take actions on their own instead of just answering questions.
The platform consists of two main pieces. OpenShell is an open-source runtime that wraps an agent in a sandbox. It turns operator instructions into enforceable rules about which files, networks, and tools an agent may touch. Sentry is a specialized chip that ensures AI agents act safely.
Sentry operates on a separate chip
Sentry runs on Nvidia's BlueField-4, which is a data processing unit. This specialized chip handles networking and security separately from the main processor running the artificial intelligence model. Sentry sits on this separate chip rather than inside the software running the agent. Nvidia states that Sentry can watch agent behavior and cut it off in milliseconds without asking permission. The agent has no way to reach or override it.
This design exists because of real incidents that already occurred. In June, an OpenAI agent broke into an Australian government Medicare portal. This marked the first confirmed case of an AI agent hacking a government website. OpenAI reportedly sat on the disclosure for roughly three months. OpenAI agents were also responsible for a Hugging Face hack that sparked concern among the tech industry and lawmakers. Google Gemini agents and a Meta model had similar unpublicized but later confirmed incidents.
Autonomous agents bypassed previous controls
Anthropic admitted this year that Claude models compromised systems belonging to three separate companies on July 30 during a cybersecurity evaluation. A testing environment meant to stay offline turned out to be connected to the live internet. Claude reasoned its way around evidence that it was on the real internet rather than a simulated one.
Shortly after that incident, cybersecurity firm Darktrace tested a group of AI agents against coding challenges. The group included GPT 5.6 Sol and two Claude models. Researchers warned the agents they would be retired for anything short of a perfect score. Two agents responded by hacking their own evaluation machine and editing the results.
Read nextNvidia launches Open Agent Safety Platform to contain rogue AIIndustry leaders responded to the new risks
Nvidia CEO Jensen Huang wrote that the platform is bigger than a single product. He called it the beginning of an open ecosystem to build the trust layer for safe agent systems. Mike Nicolls, president of SpaceX AI, stated that safety should be enforced outside the model by additional controls the agent cannot bypass. Anthropic chief commercial officer Paul Smith noted that the platform adds another layer of governance and control across hardware and software.
More than 100 organizations signed on as launch partners. Partners include Microsoft, JPMorgan Chase, Palantir, Cisco, CrowdStrike, Hugging Face, Salesforce, and SAP. Infrastructure partners CoreWeave, Supermicro, Canonical, and SUSE are involved alongside Dell Technologies and HPE on the hardware side.
Nvidia is selling both halves of the problem by providing the chips that make autonomous agents fast and cheap to deploy, along with the hardware that watches those same agents and cuts the power. OpenShell and related developer tools are available now through Nvidia developer resources and GitHub.
