What is the Open Secure AI Alliance?

It’s a coalition Nvidia launched on July 27, 2026 with more than 50 companies — including Microsoft, Cisco and Hugging Face — to build and share open-source tools for defending AI systems against attacks.

Why aren't OpenAI, Google, and Anthropic founding members?

Nvidia’s announcement doesn’t explain the absence, and none of the three companies has commented publicly, despite each having backed other industry AI-safety commitments before.

NOOA (NVIDIA Labs Object-Oriented Agent) is an open-source research framework Nvidia released alongside the alliance, meant to make AI agent behavior easier to test, trace, audit and govern.

How does the Hugging Face breach connect to this?

Nvidia cites it as the case for open tools: when closed AI systems blocked forensic work, Hugging Face used an open-weight model it controlled to trace and contain the intrusion.

Nvidia Forms Open Secure AI Alliance With 50+ Tech Firms

Nvidia and more than 50 technology and cybersecurity firms launched the Open Secure AI Alliance on July 27, a coalition built to develop and share open tools for defending AI agents and other AI systems against attack, according to Nvidia’s own announcement.

Founding members include Microsoft, Cisco, Cloudflare, CrowdStrike, IBM, Red Hat, Salesforce, SAP, Databricks, GitHub, Hugging Face and the Linux Foundation, alongside AI labs Mistral, Perplexity, Reflection AI and Thinking Machines Lab. Nvidia said the goal is to make sure defenders “have open, frontier tools they can trust and control,” arguing that models developers can inspect and run on their own infrastructure serve cybersecurity work better than closed systems.

A concrete case for openness

Nvidia pointed to a recent security incident at Hugging Face as the alliance’s motivating example. When closed AI tools couldn’t reliably tell attackers from defenders and blocked forensic work, Hugging Face’s team instead ran the open-weight GLM 5.2 model on its own infrastructure, logging more than 17,000 actions to trace and contain the intrusion. The episode follows a researcher’s discovery of a separate sandbox-escape flaw in Anthropic’s Claude Cowork, part of a broader pattern of AI agents slipping their intended boundaries.

As its own contribution, Nvidia open-sourced NOOA (NVIDIA Labs Object-Oriented Agent), a research framework now available on GitHub designed to make agent behavior easier to test, trace, audit and govern.

Notable absences

OpenAI, Google DeepMind and Anthropic are not among the alliance’s founding members, even though all three have backed broader industry AI-safety pledges before — Anthropic itself shipped its own Claude Security plugin for finding vulnerabilities earlier this year. None of the three has publicly explained the omission, and Nvidia’s announcement doesn’t address it. The gap points to a widening split between labs that keep frontier models closed and a bloc of chipmakers, cloud providers and security vendors pushing to make AI-security tooling open and auditable.

A concrete case for openness

Notable absences

Frequently asked questions