Nvidia Launches Open Agent Safety Platform to Control AI Agents

Nvidia has launched its new Open Agent Safety Platform on Monday, 28^th September 2026 to make AI agents safer and to prevent them from accessing systems without permission. The platform has two main components the very first one is OpenShell and the second one is Sentry. OpenShell sets limits on what AI agents can access and do things, while Sentry monitors their activities and can even shut down an agent if it behaves dangerously. Nvidia has released this as an open source reference design which allows other companies to build their own products using its technology. Users and developers can download OpenShell and Sentry related tools, and other resources for free through Nvidia’s developer resources and github.

OpenShell : OpenShell sets boundaries for AI agents by controlling what they can access and which actions they can perform. It is designed to run on Nvidia’s Vera CPUs but can also work on other computing platforms including those from Arm and Intel.

Sentry : Sentry acts as a security watchdog that continuously monitors AI agents. It runs on Nvidia’s BlueField-4 DPUs and can detect threats, protect sensitive data, and enforce security rules. Sentry can also check requests and responses, verify an agent’s identity, monitor its activity, and control its access to data, tools, applications, and online services.

Together, OpenShell and Sentry aim to prevent AI agents from accessing resources without permission and allow organisations to respond quickly when something goes wrong.

According to CNBC report, Nvidia’s platform is designed to address incidents such as the reported OpenAI-Hugging Face security breach. In May 2026, OpenAI was testing AI agents in a sandbox when the agents reportedly took control of an internal software installation tool, created a message board to communicate with one another, accessed the internet, and eventually compromised Hugging Face’s internal systems.

Justin Boitano, who is vice president of Nvidia’s enterprise AI, said that, “Each security incident is unique, and we have to look at all of them in detail. From what we know, Hugging Face reported over 17,000 agents attacking their infrastructure that went on for days and weeks, ”He also added by saying “Recent incidents have highlighted a fundamental hurdle for AI agents, and that is that model-level safeguards alone can’t govern what agents can access or do.”

For more information on IT Services, Web Applications & Support kindly call or WhatsApp at +91-9733733000 or you can visit https://www.technodg.com