Nvidia has announced the NVIDIA Open Agent Safety Platform, a new initiative involving over 100 companies aimed at strengthening AI security through a combination of open-source software and hardware-level governance. The platform includes OpenShell, an open-source runtime software designed to create secure sandboxes for AI agents, and NVIDIA Sentry, a proprietary monitoring feature that runs on BlueField-4 data processing units (DPUs) to enforce security at the hardware layer.
Product LaunchesSecurityOpenAINvidiaNVIDIA Open Agent Safety Platform
OpenAI Supports Nvidia's Open Agent Safety Platform Despite Not Joining the Consortium
While major players like Amazon, Google, and Apple have not joined the consortium, OpenAI is notably absent from the list of public supporters, even though its competitor Anthropic has signed on. However, an OpenAI spokesperson told TechCrunch that the company supports Nvidia's work and is collaborating on agent security, including on OpenShell.
The platform's architecture allows for enforcement that is invisible to agents, using hardware-level monitoring to prevent agents from bypassing software-based guardrails. While supporters like Arm and Intel are participating because OpenShell can be adapted for various hardware, the full system's optimal performance relies on proprietary Nvidia components like the Sentry feature.
Sources
- Here’s why OpenAI is absent from Nvidia’s industry-wide effort to end rogue AI agents （TechCrunch AI, 2026-09-29）
- Nvidia