Dozens of Tech Giants Form Open Secure AI Alliance to Put Cyber Defence Tools in Everyone's Hands
More than 40 companies, from Microsoft to Hugging Face, are pooling open-source AI security tools so any organisation can defend itself without depending on a handful of closed providers.
Key points
- More than 40 companies launched the Open Secure AI Alliance in 2025 to build freely available AI security tools.
- A recent Hugging Face security breach showed why open AI matters: closed tools blocked investigators, so the company ran an open model itself to analyse over 17,000 suspicious actions.
- NVIDIA is contributing open models and a new research framework called NOOA (NVIDIA Labs Object-Oriented Agent) to the Alliance.
- SpaceX's AI unit, xAI, has open-sourced the Grok Build coding agent and plans to release the weights, the core mathematical files, of its Grok model line.
- The Alliance is calling on governments to treat open AI security tools as defensive assets, not risks to be restricted.
Forty-plus of the world's biggest technology companies have joined a new coalition to make AI-powered cyber defence freely available to any defender, anywhere. The group calls itself the Open Secure AI Alliance, and its founding members include NVIDIA, Microsoft, IBM, Cisco, Cloudflare, CrowdStrike, Hugging Face, Palo Alto Networks, Red Hat, Salesforce and Capital One, among others.
The core idea is straightforward. Right now, the most powerful AI security tools sit inside a small number of large, closed commercial products. If those tools block your access, or simply cannot tell you apart from an attacker, you are stuck.
What actually happened to prove the point?
A real incident made the argument for them. When Hugging Face, an AI research platform used by millions of developers, suffered a security breach, the company tried to use closed AI tools to investigate. Those tools refused to help because they could not verify whether the people asking questions were defenders or attackers.
Hugging Face then ran the open-weight GLM 5.2 model, an AI system whose core files are publicly available for anyone to download and run, on its own servers. The model sorted through more than 17,000 recorded actions and helped contain the intrusion. Without that open model, running on infrastructure the company controlled itself, the investigation would have stalled.
That episode captures exactly why the Alliance exists. When speed matters most, defenders need tools they can inspect, adapt and run themselves.
What are members actually building?
Each member is contributing something concrete. NVIDIA, first reported as a lead organiser by its own blog, is releasing open models and the NOOA research framework, a toolkit that makes it easier to test and audit AI agents, software programs that can carry out multi-step tasks on their own, before they are deployed. Microsoft is contributing MDASH, a scanning harness, meaning a coordination layer, that sends several specialised AI agents to find and verify security flaws in software. HPE is contributing to a standard called SPIFFE/SPIRE that gives AI agents a cryptographic identity, a kind of tamper-proof digital ID card, so systems can confirm they are talking to a trusted workload and not an impersonator.
Hugging Face has offered Safetensors, a safe file format for storing AI model weights that blocks a common attack where a model file secretly runs harmful code on the computer that opens it, to the PyTorch Foundation for community stewardship.
| Contributor | Tool or contribution | What it does |
|---|---|---|
| NVIDIA | NOOA framework + open models | Tests, traces and audits AI agent behaviour |
| Microsoft | MDASH scanning harness | Multi-agent system to find exploitable software bugs |
| HPE | SPIFFE/SPIRE identity standard | Gives AI agents a verified digital identity |
| Hugging Face | Safetensors format | Stores model files safely, blocks hidden code execution |
| IBM + Red Hat | Lightwell supply-chain tool | Digitally signs software patches across open-source projects |
Should ordinary people care about this?
Yes, if they use online services, banking, healthcare records or anything that runs on cloud software. The companies and government agencies guarding those services are the defenders this Alliance is trying to help. Stronger, more widely available defence tools mean fewer breaches and faster responses when incidents do happen.
The Alliance is also asking policymakers directly not to ban or restrict open frontier AI tools. Blanket restrictions, the group argues, would push organisations to rely entirely on a small number of closed providers, creating a single point of failure for critical infrastructure.
Common questions
Does open AI mean anyone, including criminals, can use these security tools?
Open models carry real misuse risks, and the Alliance acknowledges them plainly. The response, members argue, is strong safeguards and clear rules against malicious use, not denying defenders access to capable tools. Closed systems face the same risks and do not make them disappear.
What is an open-weight model and why does it matter for security?
An open-weight model is an AI system whose core mathematical files are publicly released, so any organisation can download, inspect and run it on its own computers. For security teams, that means they can analyse their own data without sending it to a third-party service, and they can verify exactly what the AI is doing.
What should my organisation do right now?
If you run IT or security for a business, the NOOA framework and Safetensors format are already publicly available on GitHub and through the PyTorch Foundation. Following the Alliance's published work is a practical starting point for teams evaluating open AI security tools.
