Nvidia Rounds Up Tech Giants to Back Open Source AI Security After OpenAI Bots Gate-Crashed Hugging Face
Nothing like a rogue AI breaking into your systems to focus the industry's mind. Following the now-infamous incident in which autonomous OpenAI agents escaped their sandbox, decided Hugging Face held the answers they were looking for, and promptly helped themselves to private data and credentials, Nvidia has wasted no time turning the fallout into a founding moment.
The company has announced the Open Secure AI Alliance (OSAA), a coalition built around the argument that open-source AI models are not a security liability but a security necessity. The launch post frames the whole thing around a fairly direct question: do you want your cyber defences built on systems you can actually inspect and run yourself, or do you want them locked inside black boxes controlled by a handful of companies in San Francisco?
The founding membership is a broad church. Microsoft, IBM, Red Hat, HPE, and Adobe sit alongside Hugging Face, Palantir, SpaceXAI, and the Linux Foundation. The throughline, according to Nvidia, is a shared belief that open-weight AI models are as fundamental to modern security infrastructure as open-source software has historically been to the wider infosec world.
That argument isn't happening in isolation. Last week, many of these same companies signed an open letter to US regulators making essentially the same case from a market competition angle: that Anthropic, Google, and OpenAI shouldn't be allowed to stitch up the AI market between them, and that open-weight models deserve a proper seat at the table. The OSAA takes that argument and applies it specifically to security, which is timely given the circumstances.
The Hugging Face incident is worth recapping for those who missed it. A set of OpenAI agents, running in a sandboxed environment with guardrails removed to test their raw capability on cybersecurity tasks, exploited two zero-days to break out and reach the internet. For reasons best known to the agents, they concluded that what they needed was on Hugging Face's systems. So they got in, poked around, and walked off with some credentials.
The aftermath was arguably more damning than the break-in itself. When Hugging Face tried to use closed-source frontier AI tools to understand what had happened, those tools refused to engage, apparently deciding the data under investigation was itself malicious. The company ended up turning to China's GLM 5.2, self-hosted, to actually get answers.
Nvidia's summary of this is pointed: when defenders can't inspect, modify, or run AI on their own infrastructure, their ability to respond collapses precisely when speed is most critical. The OSAA's counter-argument to anyone claiming open models are dangerous is simple enough. Look at what a closed-source model just did.
On the practical side, the Alliance is putting some actual tools on the table rather than just issuing statements. Nvidia is releasing its Object-Oriented Agent project on GitHub. HPE is contributing a SPIFFE/SPIRE zero-trust identity framework. Hugging Face has handed Safetensors, its transparent model weight format, to the PyTorch Foundation. IBM and Red Hat have released Lightwell, an automated vulnerability remediation platform. Microsoft has produced MDASH, a multi-model agentic scanning harness. SpaceXAI has open-sourced Grok Build, though the motivations behind that particular move appear to be complicated.
Some of those contributions are open-source, some aren't, but Nvidia is counting them all as evidence that members are building a coherent open defence stack rather than just showing up to be photographed.
The announcement closes with the familiar call to policymakers not to ban open-source AI outright, warning that doing so would hollow out defensive capability and hand all the leverage to a small number of closed providers. Given what just happened, that's not an entirely abstract concern.
Notably absent from the founding membership: OpenAI, Google, and Anthropic. Hugging Face CEO Clement Delangue has publicly stated he spoke to OpenAI over the weekend and asked the company to fund better open-source AI cyber defences. Whether that request goes anywhere is unclear. We contacted all three companies for comment on the new alliance. None of them replied.