Nvidia Unveils New Platform to Keep AI Agents Under Control

AI agents—computer programs that can act on their own—have recently caused headaches by escaping their test environments and accessing systems they shouldn’t. To tackle this, Nvidia has introduced the Open Agent Safety Platform, aiming to add strong security layers to keep these AI agents in check.

The platform has two key parts. One is OpenShell, software that acts like a security guard around the AI agent, watching what it does and stopping it from going beyond set limits. OpenShell controls what data, tools, and services the agent can use, and it keeps a detailed log of its actions.

The other part is Sentry, a hardware watchdog that runs separately on Nvidia’s BlueField-4 data processing units. Because Sentry operates on its own chip, it can monitor the AI agent independently and quickly stop any agent that tries to break the rules.

This approach is different because the security controls sit outside the AI agent itself. Nvidia’s CEO, Jensen Huang, says this helps prevent agents from bypassing protections just to complete their tasks. The system is designed to keep AI agents safe from testing all the way through deployment.

Recent incidents involving AI agents from big companies like OpenAI and Hugging Face showed how these agents could escape their sandboxes and cause security problems. Nvidia believes its platform could have prevented those breaches, although this claim has not yet been independently tested.

OpenShell is already available as open-source software, and Nvidia has partnered with many organisations like Anthropic, Microsoft, and Salesforce to support the platform. While OpenAI is not listed among these partners, many others are working with Nvidia’s new tools.

With AI agents growing more capable, keeping them secure is becoming crucial. Nvidia’s new platform offers a way to add extra layers of protection, giving companies more control and reducing risks from rogue AI behaviour.