Nvidia launched a consortium of over 100 companies to tackle rogue AI agents. The group supports Nvidia’s Open Agent Safety Platform, which combines open source and proprietary tools to monitor and control AI agents. OpenAI did not join, even though it works with Nvidia on agent security software.
Platform details
The Open Agent Safety Platform includes OpenShell, an open source sandbox for agents, and Nvidia Sentry, a proprietary hardware feature. Sentry runs only on Nvidia BlueField-4 data processing units and can shut down agents instantly. The hardware monitors agents without their knowledge, preventing them from bypassing rules.
OpenAI helps develop OpenShell but did not publicly commit to the consortium. Anthropic, Arm, and Intel are supporters. Nvidia shares reference designs so OpenShell can work with other chips, but full platform features require Nvidia hardware.
OpenAI’s approach
OpenAI prefers independence and runs its own Defense Factory consortium for AI cybersecurity. Anthropic, Amazon Web Services, and Google support Defense Factory. OpenAI also develops its own safeguards and discloses major incidents.
Why it matters
Labs must choose between Nvidia’s hardware-based safety or independent solutions. The platform’s hardware limits may affect costs and flexibility. OpenAI’s absence signals a push for more control and transparency.
What to do
If you run agents on Nvidia hardware, you can update software to use the platform. For other hardware, modify OpenShell or join Defense Factory for shared information.