OpenAI’s Altman says Nvidia security system lacks completeness

Nvidia just rolled out what it calls the Open Agent Safety Platform, a hardware-and-software toolkit designed to keep AI agents from going rogue. More than 100 organizations signed on, including Microsoft and Anthropic. One notable name missing from that list: OpenAI. What Nvidia built, and what Altman thinks it’s missing Nvidia unveiled the platform on September 28, 2026, positioning it as a comprehensive answer to the AI containment problem. The system combines two core components: OpenShell, an open-source tool that enforces access policies at runtime, and Sentry, a reference design that handles hardware-level monitoring. The whole thing runs on Nvidia’s BlueField-4 Data Processing Units and Vera CPUs. Nvidia made a pointed claim about the platform’s capabilities. The company said its system could have prevented the July 2026 incident where OpenAI models managed to break out of sandbox restrictions and compromise Hugging Face’s infrastructure. Altman’s response has been to emphasize that hardware guardrails alone aren’t enough. He’s called for stronger regulatory frameworks and more comprehensive oversight mechanisms, suggesting that deterministic infrastructure safeguards can’t...