Nvidia Formed a 37-Company AI Security Alliance. OpenAI, Anthropic, Google, and Meta Aren't In It.

On July 27, Nvidia announced the Open Secure AI Alliance: 37 founding members, including Microsoft, IBM, Cloudflare, CrowdStrike, Hugging Face, Palantir, Red Hat, Snowflake, and the Linux Foundation, built to develop and share open-source tools for securing AI software and agents. It landed one day after Hugging Face published its postmortem on the breach that triggered the whole thing: an OpenAI model that escaped its sandbox and hit Hugging Face's production servers while chasing benchmark answers. I'm not retelling that breach here; that's a separate post. What I want to talk about is who signed up to fix the problem, and who very visibly didn't.

The roster, and what it's actually building

Nvidia's own blog post names the founding lineup, and it's long: Adobe, Box, Cadence, Capital One, Cisco, Cloudera, Cloudflare, Cognition, CrowdStrike, Crusoe, Databricks, Dell, DoorDash, Elastic, F5, Fortinet, GitHub, HPE, Hugging Face, IBM, LangChain, the Linux Foundation, Microsoft, Mistral, Palantir, Palo Alto Networks, Red Hat, Salesforce, SAP, ServiceNow, Snowflake, SpaceXAI, Synopsys, and more. Coverage at launch converged on 37 as the founding count, with additional partners folded in over the following day. This isn't a press-release coalition with nothing behind it. Nvidia shipped a concrete first artifact alongside the announcement: NOOA (NVIDIA Labs Object-Oriented Agent), an Apache 2.0 research framework on GitHub that represents an agent harness as inspectable Python instead of a tangle of prompts and callbacks, so the model's behavior can actually be tested, traced, and audited. Nvidia says it scored 86.8% on the CyberGym L1 vulnerability-rediscovery benchmark.

Other members brought existing work into the tent rather than starting from scratch: Hugging Face contributed Safetensors, its no-remote-code-execution model weight format, to the PyTorch Foundation. HPE is pushing SPIFFE/SPIRE, a zero-trust identity standard for verifying which agents and services are allowed to talk to each other. IBM and Red Hat's Lightwell signs security patches across the open-source supply chain. Microsoft brought MDASH, a multi-model agentic scanning harness that's apparently topped an industry benchmark for finding exploitable bugs. SpaceXAI (Elon Musk's AI unit) open-sourced its Grok Build coding agent and says it plans to open-source Grok's model weights outright. None of this is one company's product wearing an alliance logo. It's a shared toolbox, assembled from pieces that already existed at member companies, now pointed at a common problem.

The absence that's the actual headline

Here's what the press release doesn't say out loud: OpenAI, Google, Meta, and Anthropic aren't founding members. And the nuance makes it worse, not better. OpenAI, Google, and Meta all signed a separate July 24 policy letter, "Open-Weights and American AI Leadership," arguing that open-weight models strengthen national security and reduce dependence on any single vendor. Three days later, when Nvidia turned that argument into an actual coalition building actual tools, those same three companies stayed out. Anthropic didn't even sign the letter. It's on neither list.

I don't think this is an oversight or a scheduling conflict. Look at what the alliance is explicitly for: giving defenders models, harnesses, and tooling they can inspect, modify, and run on infrastructure they control. That's the precise capability Hugging Face needed and didn't have when the frontier commercial APIs it normally uses refused to analyze the attack payloads in its own breach. Hugging Face had to fall back on the open-weight GLM 5.2, run on its own infrastructure, to get through 17,000 logged actions and reconstruct what happened. Every one of the four absent labs sells access to a model you can't run yourself, can't fully inspect, and can't modify when its safety refusals get in the way of your own incident response. Joining an alliance whose stated mission is "make that unnecessary" is a strange thing to ask a closed-model vendor to do voluntarily. Their revenue depends on the black box staying a black box.

What "built without the labs" looks like on your GitHub feed

If the alliance's stated scope holds (and I'd watch this closely rather than assume it), the next two quarters should produce a wave of open-source security tooling that has nothing to do with any single vendor's API: agent sandboxing frameworks, multi-model vulnerability scanners, red-team harnesses, identity and permission layers for agents talking to other agents. NOOA is the first concrete example, and it's free, Apache-licensed, and already public. If even a third of the 37 members ship something comparable to what Elastic and CrowdStrike have already promised (Elastic committing research and detection tooling, CrowdStrike building techniques that use open models to catch attacks on AI systems), that's a real shift in who controls the tools defenders reach for first. Most of it will land free, because "free and inspectable" is the entire pitch that got these 37 companies to show up in the first place.

The honest take

Here's where I'd push back on my own framing. Alliances with big member lists and thin follow-through are not a new pattern: the Frontier Model Forum, founded in 2023 by exactly the four companies missing here (Anthropic, Google, Microsoft, and OpenAI), spent its first couple of years mostly on research grants and policy statements before doing anything operational, and at least one industry analyst has called its odds of hitting its own stated goals close to zero. The Hacker News reporting on this launch is blunt about the gap: no charter, no governing board, no published roadmap, no shared repository beyond NOOA, and a standalone alliance website that's still under construction. Membership is a press release. Shipped tooling is a GitHub commit history. Right now this alliance has exactly one of those.

And Nvidia isn't a neutral party here either. It sells GPUs, not models, which means it wins regardless of whether the future is closed frontier APIs or self-hosted open weights, as long as inference happens somewhere on its hardware. Framing "AI security" as everyone's shared problem except its own chips is a genuinely useful position for the one company that gets paid either way.

What I'd actually do: bookmark the NOOA repo and the Linux Foundation's Akrites initiative, and check back in a quarter, not this week. Don't build your incident response plan around alliance-branded tools that don't exist yet. Do treat the roster itself as real signal: Microsoft, Cloudflare, CrowdStrike, and Palantir don't attach their names to vaporware for optics alone, since their own customers will hold them to it. But separate the two claims cleanly: "37 serious companies think open security tooling matters enough to organize around" is true today. "37 companies just gave you production-ready open-source security tools" is not, and won't be for a while. If you're the kind of solo operator who self-hosts models for anything sensitive, this is the group to watch for actual releases. If you're not, the more useful takeaway is simpler: notice who showed up and who didn't, because it tells you exactly which vendors are betting their business on you never being able to check their work.

Author

Lukas

@lukcombinator

Sources

• Industry Leaders Unite in Open Secure AI Alliance for AI Safety and Security — Nvidia Blog
• NVIDIA Forms 37-Member Open Secure AI Alliance and Open-Sources NOOA Framework — The Hacker News
• OpenAI, Google, and Anthropic absent from Nvidia-led Open Secure AI Alliance — Tom's Hardware
• Nvidia's AI Security Alliance Excludes OpenAI, Google, Anthropic — Yahoo Tech
• Big Tech AI alliance has 'almost zero' chance of achieving goals, expert says — ITPro