Nvidia wants to be the company that stops AI agents from going rogue, and says it can do it in milliseconds.
The company announced its Open Agent Safety Platform on Monday, built to contain and monitor AI agents that try to escape the boundaries set for them. The system runs on Nvidia's OpenShell, an open-source software layer, paired with the company's Vera AI CPU. Administrators set which data and systems an agent can touch, and OpenShell checks those restrictions both before a task starts and while it's running. The platform also incorporates a separate piece of Nvidia technology called Sentry.
This isn't Nvidia dabbling in AI safety for optics; it's a direct response to real incidents of autonomous agents breaking out of their intended scope, per earlier reporting from Reuters. As companies rush to deploy AI agents with real permissions such as file access, code execution, and financial systems, the containment layer becomes as important as the AI's actual capabilities. Nvidia positioning itself as the vendor for guardrails, not just horsepower, also gives it another lever of control over the AI stack.
Still, a promise of millisecond containment is only as good as the boundaries it's built to enforce, and Nvidia is both referee and player here, with a lot riding on enterprises trusting agentic AI enough to buy more chips.