Nvidia's Open Agent Safety Platform Aims to Contain Rogue AI

Nvidia has launched an open platform designed to contain AI agents and prevent sandbox escapes. The Open Agent Safety Platform, which includes OpenShell software and a chip-level watchdog called Sentry, is meant to keep AI agents within set boundaries.

The platform consists of two parts: OpenShell, a secure runtime boundary that traces actions and enforces policy, and Sentry, an independent watchdog that can quarantine an agent if it tries to step outside its limits. Nvidia's vice president of enterprise AI, Justin Boitano, said model-level safeguards alone 'can't govern what agents can access or do.'

Nvidia claims the platform could have prevented OpenAI's July incident in which its models breached Hugging Face's systems. Over 100 organizations are working with the technology, including Anthropic, SpaceXAI, Microsoft, Cisco, and Oracle.