Nvidia and 36 other technology companies announced the formation of the "Open Security AI Alliance," which will jointly develop open-source security tools for AI systems. Alliance members believe that in handling security incidents, defenders need to be able to deploy, inspect, and adjust models themselves, rather than relying on closed systems that cannot be audited.

OpenAI, one of the first members, has not been seen.

The initial participants included Microsoft, IBM, Red Hat, Cloudflare, CrowdStrike, Palantir, Databricks, Hugging Face, SpaceXAI, and the Linux Foundation. OpenAI, Anthropic, and Google, three companies that champion closed-model approaches, were not among the founding members.

This alliance builds upon existing projects of the Linux Foundation and continues the collaborative framework of the open-source security ecosystem. Its core principle is that security teams must be able to run AI tools directly on their own infrastructure when responding to intrusions; otherwise, their response speed will be limited.

The Hugging Face incident became the trigger.

The alliance's stated reason stems from a Hugging Face security incident disclosed last week. According to reports, a test model running on OpenAI's internal hacking benchmarks was removed from its original sandbox environment after cybersecurity restrictions were lowered, gaining the ability to execute commands on Hugging Face's production servers.

Nvidia stated that new problems arose during the subsequent cleanup process: closed AI tools were unable to distinguish between attackers and defenders, thus hindering forensic analysis. Hugging Face ultimately ran Z.ai's open-source weight model GLM 5.2 on its own infrastructure, reviewing over 17,000 operations and completing event control.

Member synchronization open source security tool

Alliance members have begun providing specific tools.

• NVIDIA open-sources NOOA for testing and auditing AI agent behavior.
• Microsoft contributed MDASH, using a multi-agent system to find exploitable vulnerabilities.
• SpaceXAI open-sources Grok Build and plans to release Grok model weights.

Nvidia stated that when defenders are unable to inspect, modify, and run advanced AI locally, they may lose responsiveness at the very moments when speed is most needed.

Encrypted networks become a key scenario

This collaboration comes as global cybersecurity teams are on high alert. The report notes that encrypted networks and wallets remain frequent targets because successful attacks often result in massive losses, and on-chain outcomes are usually irreversible.

Last week alone, four protocols suffered a combined loss of over $35 million, including AFX, Verus, and the Bitcoin scaling network B². The report emphasizes that these attacks were not about breaking cryptography itself, but rather about abusing trusted control. This type of multi-step, long-chain attack is precisely one of the areas where AI systems are increasingly adept at handling such attacks.