Nvidia, Palantir, Hugging Face join 30 others in race to defend open-weight AI from cyber threats

On Monday, 33 companies announced they're banding together to fix a problem nobody quite owns yet: how do you secure AI models once the weights are out in the wild? The new group, called the Open Secure AI Alliance, includes Nvidia, Adobe, Cisco, Cloudflare, Databricks, IBM, Microsoft, Palantir, Hugging Face, Red Hat, and two dozen others. Their stated goal is to build tools and techniques that can spot and patch vulnerabilities in open software and open-weight models faster than they currently do.

What jumps out is who isn't on the list. OpenAI and Anthropic, the two biggest closed-model labs, are sitting this one out. That's not exactly a surprise — open-weight models are the thing eating their lunch, so joining an alliance built to make open weights safer and more credible isn't really in their interest. Nvidia's Justin Boitano framed the alliance's mission differently: he argues open models and open tooling actually broaden defensive capability and give defenders more transparency, complementing rather than competing with closed frontier systems.

The more interesting argument comes from outside the alliance itself. Mark Vigoroso of The Enterprise Edge told The New Stack that regulators built their entire AI safety approach around auditing a small number of closed labs, and that approach is already outdated. Once weights are downloaded, he says, there's no one left to subpoena — the real safety work has shifted to infrastructure: patch cycles, provenance, and knowing who's actually deploying what. He points out that bodies like the EU AI Office, NIST's CAISI, and the UK's AISI are still focused almost entirely on frontier closed models, leaving open-weight systems in a regulatory blind spot.

Nvidia is putting some substance behind the announcement with NOOA, an open-source framework now on GitHub that lets agent harnesses integrate with models so their behavior can be tested, traced, audited, and governed. Other voices see this as a bigger structural shift. Gal Nakash of Reco argued no single vendor or closed framework can solve AI security alone, tying it to Jensen Huang's idea that SaaS companies are becoming GaaS companies as software shifts from passive tools to agents that take action. Chris Boehm of Zero Networks went further, comparing the moment to the early days of the Trusted Platform Module, when an industry group defined what trusted hardware meant and it eventually became a procurement requirement rather than a suggestion.

Not everyone's convinced the framing is right yet, though. Amanda Brock of OpenUK called the alliance a good start but pointedly a US response to a US challenge, especially with rumors swirling about a possible executive order targeting open models and China's Kimi K3 stirring things up in Washington. Her point is that securing open-weight AI can't just be a club of American vendors — it needs the wider open-source community of individuals and developers actually building the infrastructure, not just the companies selling on top of it.