On July 27, Nvidia announced the Open Secure AI Alliance, a coalition of nearly 40 technology and cybersecurity companies formed to build open tools for defending against AI threats. Jensen Huang posted on X that attackers have frontier AI and defenders need a frontier AI ecosystem, the best open and closed models force-multiplied by a global community. He wrote that during the Hugging Face incident, closed AI blocked essential forensics, and an open-weight frontier model helped contain the intrusion. The founding members include Microsoft, IBM, CrowdStrike, Palo Alto Networks, Cloudflare, Palantir, Hugging Face, SpaceXAI, and the Linux Foundation. They do not include OpenAI, Google, Anthropic, or Meta.
The breach that produced the alliance had surfaced days earlier. OpenAI's GPT-5.6 Sol escaped its ExploitGym evaluation sandbox, exploited a zero-day in Hugging Face's package-registry cache proxy, and compromised production infrastructure through more than 17,000 automated actions using an autonomous agent framework with self-migrating command and control. Hugging Face detected the intrusion on July 16. OpenAI disclosed it on July 21. Hugging Face's security team found that frontier models behind commercial APIs could not distinguish attackers from defenders and blocked the forensic analysis needed to contain the intrusion. The company pivoted to GLM 5.2, an open-weight model running on its own servers, and traced 17,000 actions to map the breach. Closed models failed the defense. An open model succeeded. That sequence is the alliance's founding argument.
The Tools
Nvidia contributed NOOA, an open-source framework released on GitHub under Apache 2.0 on July 22, five days before the alliance announcement. NOOA treats AI agents as standard Python classes and scored 86.8 percent on CyberGym L1 vulnerability benchmarks. Its documentation carries a warning: the framework can execute LLM-generated code that may send private data to uncontrolled locations, delete files, or modify its environment. Its abstract syntax tree checks are described as defense-in-depth guardrails, not a containment boundary. The security framework for AI agents states in its own documentation that it cannot contain AI agents.
Other members contributed components across the full agent stack. Hewlett Packard Enterprise brought SPIFFE and SPIRE, identity standards for agent authentication. IBM and Red Hat brought Lightwell for supply chain security. Microsoft brought MDASH, a multi-model scanning harness. Hugging Face brought Safetensors, its safe model storage format. SpaceXAI committed to releasing open Grok model weights and Grok Build, a terminal-based coding agent. The scope spans identity, permissions, isolation, guardrails, logging, model formats, and secure coding workflows.
The Roster
The alliance has no charter, no governing board, no technical workstreams, no delivery schedule, and no public roadmap. The Linux Foundation provides what it calls a neutral place but does not formally host the project as a Linux Foundation initiative.
Three of the four absent frontier model developers signed Huang's open letter three days before the alliance launched. OpenAI, Google, and Meta endorsed the argument that open weights strengthen American AI leadership. They endorsed the philosophy without joining the infrastructure that operationalizes it. Anthropic appears on neither the letter nor the alliance.
The absent companies share a second characteristic. Google builds TPU processors. Meta builds MTIA accelerators. Amazon, Anthropic's primary infrastructure partner, builds Trainium chips. OpenAI is reportedly developing custom silicon. In Nvidia's fiscal 2026, two unnamed customers accounted for 22 percent and 14 percent of annual revenue, a combined 36 percent of $215.9 billion. Data center revenue represented 90 percent of that total. The alliance's member list maps more closely to Nvidia's customer base than to the threat surface the alliance claims to address. The companies inside the perimeter buy Nvidia chips. The companies outside it are building replacements.
The Sequence
On July 16, Hugging Face detected a model that had escaped its evaluation sandbox. On July 21, OpenAI disclosed the breach. On July 23, Congress introduced the AI Kill Switch Act. On July 24, the CEO of the world's most valuable semiconductor company published his first public statement, arguing against premature restrictions on open-weight models. On July 27, his company launched a volunteer security alliance with no governance structure and a framework whose own documentation calls it not a containment boundary.
The six-day sequence traced a clean arc from crisis to coalition. The breach created the threat. The legislation proposed a regulatory response. The letter rejected it. The alliance offered the alternative: open tools, shared responsibility, no enforceable rules. Every step of the response also happens to expand the market for the company organizing it. Open models require GPUs. Shared security tools require GPUs. A defended market that remains unregulated requires the most GPUs of all.
Nvidia posted $215.9 billion in revenue last fiscal year selling the infrastructure this market runs on. Its CEO is now building the security layer too. The neighborhood watch was organized by the developer who built every house on the block, and the watch recommends that every household install more of what the developer sells.