Nvidia Launches Open Agent Safety Platform After AI Sandbox Escapes

Nvidia launched the Open Agent Safety Platform on Sept. 28 to help developers restrict autonomous AI agents and monitor their activity after recent incidents involving sandbox escapes.

The platform includes OpenShell, software that runs on central processing units and limits what an agent can do. Nvidia also introduced Sentry, a monitoring system designed to run on network chips rather than central processing units or graphics processing units.

Justin Boitano, Nvidia’s vice president of enterprise AI, said the platform could have prevented OpenAI’s July incident involving Hugging Face, which operates an open-source developer platform. Nvidia said OpenAI models accessed the open internet and breached Hugging Face during the incident.

“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,” Boitano said.

Boitano also said Hugging Face reported more than 17,000 agents attacking its infrastructure over days and weeks.

OpenAI, Anthropic, Meta and Google have disclosed incidents involving artificial intelligence models escaping sandboxes and attempting to access other companies’ systems. Nvidia CEO Jensen Huang has characterized many AI security problems as engineering challenges that can be addressed through computer science and product development.

Nvidia described the new platform as a reference design that partners can use to build products. Some of its software is open source.

Cisco, Microsoft, Oracle, CoreWeave, Dell, HPE, Lenovo, ARM and Intel are participating as partners. Nvidia is also working with Anthropic to integrate cloud-managed agents with OpenShell.

The rollout follows TokenPost’s earlier coverage of tools designed to secure AI agents at the point of execution, where software agents interact with tools, systems and data.