Nvidia Assembles Fifty-Plus Firms Behind Open AI Security — Without OpenAI, Anthropic, Google or Meta

Nvidia launched the Open Secure AI Alliance on 27 July, a coalition built around the argument that cyber defenders need open frontier AI models alongside closed ones, and should not have to depend on a handful of proprietary systems.

The founding roster spans cloud, security, enterprise software and open-source foundations: Microsoft, IBM, Red Hat, Cisco, Cloudflare, CrowdStrike, Palo Alto Networks, Dell, HPE, Hugging Face, the Linux Foundation, Salesforce, SAP, ServiceNow, Siemens, Snowflake and Zscaler among them.

Who is in, and who is not

OpenAI, Anthropic, Google and Meta are absent. Those four run the closed frontier models the alliance says defenders should be able to use alongside open ones, which makes their absence the most-discussed feature of the launch. Some of them signed a separate industry letter on 24 July.

Early coverage placed the count at 27, 37, 40-plus and 44. Nvidia's own announcement does not give a total but names 52 organisations as inaugural partners — the only figure traceable to a primary list, so that is the one we used.

Regionally the roster leans North American and European, with NAVER and SK Telecom from South Korea and G42 from the UAE. No Southeast Asian organisation appears on the founding list.

What is actually being released

The alliance is shipping code rather than only a position, though several contributions predate the coalition and are being brought under its umbrella.

NOOA — the NVIDIA Labs Object-Oriented Agent — takes the approach of writing agent harnesses as ordinary Python classes, where methods left with ellipsis bodies are completed by a model and everything else stays deterministic code. The design goal is auditability: you can see which parts of an agent's behaviour are decided by a model and which are not.

The argument being made

Nvidia's framing is that defenders need both kinds of system. Its statement says defenders need frontier closed models and frontier open models working together, so they can pick the right system for the job and have transparency, adaptation and sovereign control available wherever security demands it.

The founding position goes further, arguing that blanket restrictions on open frontier AI would weaken defensive capacity and risk concentrating power among closed providers. Nvidia also concedes the counter-argument directly, acknowledging that open models can be misused while contending those risks are not unique to open systems and must be managed wherever advanced AI runs.

The alliance builds on existing open-source security work, including the Linux Foundation's Akrites project and OpenSSF.

What it is built on

The alliance is not starting from nothing. It positions itself as an extension of existing open-source security work, naming the Linux Foundation's Akrites project and OpenSSF — the foundation's security arm, best known for the supply-chain tooling that followed the Log4j period.

That lineage matters for reading the contribution list. Safetensors is the clearest case: Hugging Face released it to replace Python pickle serialisation for model weights, a format that executes arbitrary code on load and made downloading a model a code-execution risk. It became a default across the ecosystem on the strength of solving a real problem, with no consortium behind it.

NOOA is the piece to judge the alliance on, because it is the only contribution introduced for the launch. The word harness does quiet work here: it is the scaffolding that decides what a model can call, what it can see, and where its output goes. Most agent security failures sit in the harness, not the model. A framework that makes that boundary inspectable in source is a more substantive offering than the position paper beside it.

The timing

The coalition tied its case explicitly to the July intrusion at Hugging Face, where an autonomous agent system compromised parts of the company's production infrastructure. We covered that episode separately: OpenAI later said the agents were its own pre-release models, running with cyber refusals reduced for an internal evaluation.

That case supplies an awkwardly good example for both sides of the open-versus-closed argument. The attacking models were closed and pre-release. The model Hugging Face's responders used to investigate was open-weight, chosen because commercial APIs refused the analysis.

What it does not settle

An alliance is a statement of intent, and this one arrives without a governance structure, funding commitment or certification scheme announced. What it has is a roster, a licence and a first repository.

Whether it becomes infrastructure or a logo wall turns on whether the contributions are maintained after launch week. Safetensors is a reasonable precedent — it became a de facto standard on merit rather than through a consortium — and it is also a reminder that the successful open security tools on this list succeeded before any alliance existed.

The caveats

• Member counts vary by source. We used the 52 organisations named in Nvidia's own announcement; published figures ranged from 27 to 44 in the first day of coverage.
• Several contributions are pre-existing. Safetensors and MDASH were released before the alliance; only NOOA is new to the launch.
• Nvidia has a commercial interest in a large, plural ecosystem of models running on its hardware. That does not make the argument wrong, but it is not a neutral party making it.
• No governance or funding detail has been published.

Key takeaways

• The Open Secure AI Alliance launched 27 July with 52 organisations named by Nvidia as inaugural partners.
• OpenAI, Anthropic, Google and Meta are not founding members.
• NOOA is the first new contribution, published under Apache 2.0 and aimed at making agent behaviour auditable.
• The case was tied to the Hugging Face intrusion, which cuts both ways on open versus closed.
• No governance, funding or certification structure has been announced.