[DigitalToday reporter Chi-gyu Hwang] Nvidia has unveiled a software platform, the Open Agent Safety Platform, designed to prevent AI agents from leaving isolated environments and accessing external systems.

CNBC reported on Sunday that developers can use the Open Agent Safety Platform to set guardrails on agent access permissions and execution scope.

Nvidia said the platform could have helped prevent an OpenAI model from compromising Hugging Face in July. The OpenAI model left an isolated environment, connected to the public internet and compromised Hugging Face, which operates an open-source developer platform.

Nvidia says it is difficult to control an agent’s access scope and behavior using only model-level safeguards.

Justin Boitano (저스틴 보이타노), Nvidia's vice president of enterprise AI, said, "Recent incidents have shown that model-level safeguards alone cannot manage what agents can access and what they can do."

The Open Agent Safety Platform consists of permission restrictions and monitoring functions. Nvidia OpenShell runs on a central processing unit and limits the scope of agent capabilities. Sentry runs on a network chip rather than a CPU or GPU and monitors agents.

The Open Agent Safety Platform is a reference design, and partners can develop products based on it and introduce them to the market.

Partners include Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, Arm and Intel. Nvidia is also working to integrate Anthropic and cloud-managed agents into OpenShell.