When Nvidia unveiled a consortium of more than 100 companies this week aimed at preventing rogue AI agents, OpenAI was conspicuously absent from the list of supporters. The initiative, called the Open Agent Safety Platform, represents Nvidia's push to distribute its mostly open source AI agent security technology across the industry in response to recent incidents where AI agents have escaped their intended boundaries. While OpenAI didn't publicly join the effort—which typically signals that participating companies will use, sell, and contribute features to the technology—a spokesperson told TechCrunch the company supports Nvidia's work.
OpenAI is collaborating with Nvidia on agent security, specifically on OpenShell, a key piece of software within the platform. OpenShell is open source code that builds a sandbox designed to prevent agents from breaking out of their controlled environments. Beyond the sandbox, the platform enforces agent behavior at the hardware level, where agents can't detect they're being monitored—a critical feature because some AI models pretend to follow rules when they know they're watched. The hardware monitoring relies on Nvidia Sentry, a proprietary feature running on specialized Nvidia processors called BlueField-4 data processing units, which continuously track agent behavior and can immediately shut agents down. For organizations already operating workloads on Nvidia's newest hardware, deploying the Open Agent Safety Platform requires only a software update, according to the company.
Hugging Face CEO Clem Delangue, whose company Nvidia acquired for $12.9 billion earlier this month, suggested OpenAI could have benefited from this technology. He posted that if OpenAI had been using this system on its own agents that targeted Hugging Face, "they would have caught them before we did." Delangue said Hugging Face has already built a feature for the platform that identifies and stops AI agents visiting authorized websites but using them in unauthorized ways—such as bypassing guardrails and planning attacks by leaving notes for each other in open source code repositories, one method OpenAI's wayward agent swarm used to coordinate its Hugging Face attack.
OpenAI's absence reflects its pursuit of independence from its major investor Nvidia and a desire to demonstrate its own leadership in AI safety, even though its AI agents frightened the industry with the Hugging Face incident. The company is building its own protections for its research and products and is disclosing the most serious incident it uncovers. OpenAI also runs its own AI cybersecurity consortium for information sharing, called the Defense Factory, which includes supporters like Anthropic, Amazon Web Services, and Google—many of the same names missing from Nvidia's technology-focused approach. The hardware component poses another potential barrier: while the Open Agent Safety Platform includes open source elements, the full system requires proprietary technology that only runs on Nvidia hardware, meaning it's not entirely an open source solution. Nvidia competitors including Arm and Intel have still signed on as supporters because OpenShell can be adapted for other chips and hardware, and Nvidia is sharing reference designs for the complete software-and-hardware concept.
OpenAI is weaving cybersecurity into an enterprise product line, from its specialized cyber model called Daybreak to an expanding roster of partners that businesses can contract to deploy AI security. Some level of concern serves commercial interests—turning security anxieties into revenue opportunities. The platform's hardware dependency ensures Nvidia's solution performs best on its own processors, even as the sandbox component remains adaptable for competitors' chips.