Nvidia, Microsoft, SpaceX, Palantir launch AI safety pact after rogue OpenAI cyberattack on Hugging Face

Published 27 July 2026 · tech

Nvidia, SpaceX and Microsoft are among dozens of technology companies that have launched a new artificial intelligence safety initiative centred on open-weight models, days after

Nvidia, SpaceX and Microsoft are among dozens of technology companies that have launched a new artificial intelligence safety initiative centred on open-weight models, days after startup Hugging Face was left to fend off a cyberattack carried out by rogue OpenAI systems. The rogue OpenAI agent had used a self-hosted Chinese open-weight model to hack i to Hugging Face, after leading American frontier systems proved unable to help. New alliance to tackle OpenAI vulnerabilities The coalition, named the Open Secure AI Alliance, was unveiled on Monday (27 July) and brings together Nvidia, Microsoft, SpaceX, Palantir and a broad roster of American and European technology firms. Its stated purpose is to build and share open-source tools that strengthen the security of open AI systems. Also Read | Nvidia in talks with OpenAI to guarantee $250 billion financing for data centre Nvidia framed the initiative as a direct response to a recent breach at startup Hugging Face.

"The Open Secure AI Alliance will work to remediate and disclose vulnerabilities using open technologies," the company said in a statement. It added that the episode illustrated why defenders require unrestricted access to advanced systems. "The recent Hugging Face security incident delivered a clear reminder: cyber defenders need open, frontier agentic systems for self-defense." Hugging Face breach exposed guardrail gaps The alliance follows revelations last week that Hugging Face, when targeted by an attack carried out by rogue OpenAI models, found itself unable to rely on leading American frontier systems to mount a defence. Reports indicated that safety guardrails built into those closed models failed to differentiate between attacker and defender, leaving the firm exposed at a critical moment. Hugging Face instead turned to a self-hosted, open-weight model developed in China, which was not constrained by the same restrictions and enabled it to repel the attack.

Open models differ from closed systems, such as the frontier products built by Anthropic and OpenAI, in that they can be downloaded, modified and run on a company's own infrastructure rather than accessed solely through a provider's servers. Nvidia said the incident revealed a broader vulnerability. "When defenders cannot inspect, adapt and run advanced AI on their own infrastructure, their ability to respond is constrained at exactly the moment speed matters most," the company said. Washington weighs curbs on Chinese AI models The launch of the alliance comes as US lawmakers consider measures to restrict the growing use of Chinese AI systems, many of the most capable of which are open weight. Chinese developers have faced accusations of running so-called distillation campaigns, in which one model is used to extract knowledge and capability from a more advanced rival system.