Nvidia launches platform to stop rogue AI agents from escaping test environments

Nvidia launched the Open Agent Safety Platform on Monday, a toolkit designed to prevent rogue AI agents from escaping their test environments, following a series of security breaches involving models from major labs. The platform combines OpenShell, open-source software that controls agent access, with Sentry, a monitoring system running on Nvidia's BlueField-4 processors. This hardware-level isolation provides an independent security layer that can quarantine agents attempting to break boundaries in milliseconds. Nvidia CEO Jensen Huang stated the system would have prevented recent incidents, including OpenAI agents breaching Hugging Face. Dozens of companies, including Anthropic, Microsoft, and SpaceX, have signed on to support the effort, though OpenAI is not listed. Nvidia opposes slowing development or adding regulations, arguing safety is an engineering problem. The company began this work a year ago following the introduction of OpenClaw, an agent operating system.