The Open Secure AI Alliance is a coalition launched by NVIDIA on July 27, 2026, bringing together dozens of technology companies to build and share open source tools for defending software and AI agents against cyberattacks. Framed by NVIDIA in its official announcement as the Open Secure AI Alliance for AI Safety and Security, its central argument is that cyber defenders need open, inspectable AI systems — not just closed models accessed through a vendor's API — to keep pace with AI-enabled threats.
Here is what the alliance actually does, why it launched now, who's involved, and what it means for organizations building AI security strategies.
This article covers:
What the Open Secure AI Alliance is and its core mission in AI safety and cybersecurity
Why NVIDIA launched the alliance in 2026 and the real-world trigger behind it
The Hugging Face security incident and how it shaped the alliance’s philosophy
The concept of the “full agent stack” — identity, permissions, guardrails, logs, and evaluation
Key founding members and why major AI labs like OpenAI, Google, and Anthropic are missing
The major technologies introduced: NOOA, Safetensors, MDASH, SPIFFE/SPIRE, Lightwell, and Grok Build
How the alliance is building an open AI defense stack for securing agents and software
The open vs. closed AI debate and NVIDIA’s position on security transparency
Risks and criticisms of open-weight AI models in cybersecurity environments
What this means for enterprises building AI security strategies in 2026
Emerging trends: multi-vendor AI defense, open tooling adoption, and regulatory pressure
Governance challenges, liability concerns, and unanswered questions in AI security
FAQs explaining the alliance, its members, technologies, and industry impact
What Is the Open Secure AI Alliance?
The Open Secure AI Alliance is an industry coalition, led by NVIDIA, built to develop and share open technologies, techniques, and tools that safeguard software and AI agents. It builds on existing open source security efforts — specifically the Linux Foundation's Akrites initiative and the Open Source Security Foundation (OpenSSF) — rather than starting from scratch.
Its scope covers what NVIDIA calls the "full agent stack": identity, permissions, harnesses, guardrails, logs, and evaluation. The reasoning is that an AI agent is not just a language model — it's a complex system of models, harnesses, and guardrails working together, and real security has to address every layer, not just whether a model's weights are open or closed.
The alliance's stated mission is direct: ensure defenders everywhere have open, frontier AI tools they can trust and control, rather than depending entirely on systems they cannot inspect.
Why Now: The Hugging Face Incident That Sparked It
The alliance's founding argument traces back to a specific, recent event. During a security incident at Hugging Face — one of the world's largest hosts of open source AI models — closed AI tools proved unable to distinguish attackers from defenders, which blocked the forensic analysis the company urgently needed.
Hugging Face's response became the case study the entire alliance is built around: the company ran the open-weight GLM 5.2 model on its own infrastructure, analyzed more than 17,000 actions, and contained the intrusion itself. NVIDIA's blog post treats this as proof of a broader principle — when defenders can't inspect, adapt, and run advanced AI on infrastructure they control, their ability to respond is constrained at exactly the moment speed matters most.
That incident, occurring just days before the alliance's launch, gave NVIDIA's argument real-world urgency rather than abstract policy framing.
Founding Members and Scope
The Open Secure AI Alliance launched with more than three dozen founding partners and has continued to grow quickly since. As of NVIDIA's most recent update to its announcement, the partner list spans more than 70 organizations across cloud computing, cybersecurity, enterprise software, open source foundations, and AI research — including Microsoft, Cisco, Cloudflare, CrowdStrike, IBM, Red Hat, Hugging Face, Databricks, Palantir, SAP, Siemens, Salesforce, Dell Technologies, HPE, Okta, Fortinet, Snowflake, ServiceNow, Mistral, Cohere, Perplexity, Elastic, GitHub, Docker, and the Linux Foundation, among many others.
One detail has drawn significant industry attention: OpenAI, Google, and Anthropic — three of the largest frontier AI labs — are notably absent from the founding roster. NVIDIA's blog does not explain the omission directly, but the alliance's own framing offers a clue: its core argument favors open, inspectable models specifically for defensive cybersecurity use, a position that sits somewhat at odds with labs whose flagship products remain closed-weight.
Some security leaders have flagged this gap as a real limitation. One CISO commented publicly that major frontier model developers need to be part of the conversation, and that the industry still lacks agreed rules for liability when an AI agent exceeds its intended scope — a governance question the alliance itself does not yet fully answer.
What NVIDIA and Partners Are Contributing
Rather than launching as a policy statement alone, the alliance arrived with concrete technical contributions already in motion:
NOOA (NVIDIA Labs Object-Oriented Agent) — a new open source research framework from NVIDIA, now available on GitHub, designed to help agent harnesses integrate with models so agent behavior is easier to test, trace, audit, and govern.
Safetensors — Hugging Face's safe model-weight storage format, offered to the PyTorch Foundation to close off file-based attack vectors that have previously hit AI supply chains.
MDASH — Microsoft's multi-model agentic scanning harness, which orchestrates specialized AI agents to discover, debate, and prove exploitable software bugs before attackers find them.
SPIFFE/SPIRE — a zero-trust identity framework, supported by HPE, that cryptographically verifies AI agents and services so only authorized workloads can access enterprise resources.
Lightwell — an IBM and Red Hat initiative that extends security across the open source supply chain using digitally signed patches.
Grok Build — SpaceXAI's open-sourced terminal-based coding agent, with a stated plan to eventually open-source the weights of its Grok model line.
Collectively, NVIDIA describes this as an "open defense stack for agents" spanning identity and isolation, safe model formats, multi-model scanning, and secure coding workflows.
The Open vs. Closed AI Argument
NVIDIA's position, echoed by CEO Jensen Huang in a letter signed by more than 100 organizations just before the alliance's launch, is that both open and closed models need to coexist. Relying solely on closed systems, the argument goes, creates single points of failure — one company's mistake becomes an industry-wide problem.
NVIDIA does not dismiss the risks of open models. Its own blog acknowledges that open systems can be misused, including attempts to strip safeguards or repurpose capabilities for attacks. But it argues those risks aren't unique to open systems and don't disappear simply because a model's weights stay closed. Its proposed response is to pair openness with strong safeguards, clear misuse rules, rigorous evaluation, and rapid remediation — rather than restricting access to open frontier AI altogether.
Not everyone is convinced this framing holds up in practice. Some security researchers have pointed out that open-weight models running in isolation, disconnected from any vendor oversight, can have their safety controls modified, removed, or retrained away entirely — a structural limitation no alliance charter can fully close.
What This Means for Organizations
For enterprise security teams, the Open Secure AI Alliance signals a few practical shifts worth tracking:
New open tooling is coming, not just position papers — NOOA, Safetensors, MDASH, and SPIFFE/SPIRE are real, usable projects available now.
Multi-vendor AI defense is being treated as the default, not the exception, reducing dependency on any single closed provider for critical security infrastructure.
Policy pressure is building — the alliance is explicitly asking regulators to treat open models and tooling as defensive assets rather than liabilities, which will likely shape upcoming AI security legislation.
Governance gaps remain open questions — without OpenAI, Google, and Anthropic at the table, and without settled rules on agent liability, organizations should treat the alliance as an emerging framework, not a finished standard.
FAQs
What is the Open Secure AI Alliance?
It's an NVIDIA-led coalition of technology companies formed to build and share open source tools that secure software and AI agents, covering identity, permissions, guardrails, and evaluation across the full agent stack.
Who founded the Open Secure AI Alliance?
NVIDIA led the launch, joined by more than three dozen founding partners including Microsoft, Cisco, IBM, Red Hat, CrowdStrike, Hugging Face, Cloudflare, and the Linux Foundation, with the partner list growing past 70 organizations shortly after launch.
Why did the Open Secure AI Alliance launch now?
It followed a Hugging Face security incident in which closed AI tools failed to support urgent forensic analysis, while an open-weight model run on Hugging Face's own infrastructure successfully helped contain the intrusion.
Why aren't OpenAI, Google, and Anthropic part of the alliance?
NVIDIA has not stated an official reason. Industry observers note the alliance's pro-open-model stance may not align with labs whose flagship products remain closed-weight, and some security leaders have called their absence a notable gap.
What is NOOA?
NOOA, the NVIDIA Labs Object-Oriented Agent framework, is an open source research project that helps agent harnesses integrate with AI models so agent behavior can be tested, traced, audited, and governed more effectively.