OpenAI Bypasses Nvidia’s Agent Security Alliance

Nvidia’s Open Agent Safety Platform aims to standardize security by embedding monitoring tools directly into the hardware layer. The initiative centers on OpenShell, a sandbox designed to isolate AI agents, and Nvidia Sentry, a proprietary feature requiring BlueField-4 data processing units. By tying security to its own hardware, Nvidia creates a powerful, albeit closed, ecosystem. This strategy offers a robust defense against rogue behavior, yet it forces participants into a reliance on Nvidia’s infrastructure that some industry leaders are clearly hesitant to accept.

OpenAI’s decision to remain outside this alliance appears to be a calculated move toward independence. The company is currently scaling its own security initiatives, most notably the Defense Factory, a consortium that includes Amazon Web Services and Google. By developing internal safeguards—such as its cyber-focused model, Daybreak—OpenAI is positioning its safety tools as a commercial enterprise product. This approach allows the company to retain control over its intellectual property while avoiding the hardware constraints inherent in Nvidia’s model. Although Hugging Face CEO Clem Delangue noted that OpenAI could have prevented past agent attacks using Nvidia’s tools, OpenAI seems determined to define the standards of AI security on its own terms.