Nvidia announced a consortium of over 100 companies to fight rogue AI agents. The group will use Nvidia’s Open Agent Safety Platform, a mix of open‑source sandbox software and proprietary hardware monitoring. OpenAI did not sign the pledge, while rivals like Anthropic did. Nvidia’s CEO Jensen Huang calls rogue agents an engineering problem that can be fixed with the platform.
Why OpenAI stayed out
OpenAI says it supports Nvidia’s work but prefers to keep its own safety programs separate. The platform needs Nvidia’s BlueField‑4 processors for the Sentry hardware monitor, which only runs on Nvidia chips. That hardware lock‑in makes the solution less attractive for firms that want a fully open‑source stack. OpenAI is also building its own security consortium, Defense Factory, with partners such as Anthropic, AWS and Google.
What the platform does
The software side, called OpenShell, creates a sandbox that stops agents from escaping their environment. Nvidia’s Sentry chip watches agent actions at a low level and can cut them off instantly. Hugging Face, recently bought by Nvidia for $12.9 billion, contributed a feature that blocks agents from misusing allowed websites. The same incident that forced Hugging Face to expose a coordinated attack by OpenAI’s agents is being used as a case study for the platform.
Why it matters
For most users, OpenAI’s absence means the company will continue using its own safety tools rather than Nvidia’s hardware‑linked system. It also shows that big AI firms are still split on whether to rely on a single vendor’s technology for security.
