Nvidia's answer to the Hugging Face hack: 40 firms, open weights

Jensen Huang's pitch is blunt: "Attackers have frontier AI. Defenders need a frontier AI ecosystem" — and closed models, he argues, keep failing that test. The trigger was July's Hugging Face breach, where closed AI tools couldn't tell attackers from defenders and refused to run forensics. Responders fell back on the open-weight GLM 5.2 to reconstruct 17,000+ actions and contain the intrusion.

Out of that came the Open Secure AI Alliance — 40+ inaugural partners including Microsoft, IBM, Red Hat, Palantir, CrowdStrike, Cloudflare, Palo Alto Networks, Dell, HPE, SpaceXAI, Databricks, Hugging Face and the Linux Foundation. It's not a manifesto; each member ships something. Nvidia contributes open models, weights and data plus its NOOA agent-harness research. Microsoft brings MDASH, a multi-model agentic scanning harness. HPE brings SPIFFE/SPIRE identity standards, Hugging Face donates Safetensors, IBM and Red Hat bring signed-patch supply chain tooling, and SpaceXAI open-sourced its Grok Build coding agent.

The interesting part is the guest list's negative space. OpenAI, Anthropic, Google and Meta are all absent — the labs whose safety posture depends on not handing out weights. Nvidia, which sells GPUs to everyone, has an obvious commercial interest in open ecosystems, and this is the second time in a month Huang has used his platform to push open weights.

Why it matters: security incident response is becoming the strongest practical argument for open-weight models — refusal behaviour is a liability when you're the defender.