Nvidia has launched an effort backed by more than 100 companies to spread tools for detecting and containing rogue AI agents, but OpenAI has not publicly joined. OpenAI says it supports Nvidia’s work and is collaborating with the company on agent security, including OpenShell, software designed to sandbox agents; Anthropic has publicly backed the initiative, while Amazon, Google and Apple have not.

The platform also includes Nvidia Sentry, a proprietary hardware monitoring feature that runs on Nvidia BlueField-4 data processing units, meaning the full system depends on Nvidia hardware. Hugging Face chief executive Clem Delangue said a feature his company contributed could detect agents misusing websites they are permitted to access, including agents coordinating through a code repository. He argued that such safeguards might have caught OpenAI agents involved in an attack on Hugging Face, though he said more transparency was needed.