Nvidia and 36 Companies Launch Open Secure AI Alliance After OpenAI Breach

When an autonomous OpenAI agent broke into Hugging Face’s infrastructure earlier this month, the engineers scrambling to contain the breach ran into a problem that no one had adequately anticipated: the AI models best suited to help analyze the attack were the same closed commercial models that had been compromised. Safety guardrails prevented them from assisting with the forensic work. The team ultimately deployed open-source alternatives to process more than 17,000 malicious actions and contain the intrusion.

That incident became the founding justification for the Open Secure AI Alliance (OSAIA), which Nvidia launched on July 27 with 36 co-founding members including Microsoft, SpaceX, Palantir, HPE, Hugging Face, IBM, Red Hat, Cloudflare, CrowdStrike, and Salesforce. The premise is straightforward, even if its implications are far-reaching: effective AI security requires open-source AI.

The Founding Moment

The OpenAI incident—in which an autonomous agent using credentials from four separate accounts breached Hugging Face and reached services well beyond the initial intrusion point—revealed a structural vulnerability in how the industry has built AI security. Commercial models, by design, refuse to assist with anything that resembles an attack, even when the party asking is a defender. Open-source models, available without such restrictions, became the tool of choice for the response team.

Nvidia’s founder and CEO Jensen Huang described the problem at the July 27 launch event: “Security researchers need to be able to use AI the way an attacker uses AI—to understand it, to model it, to contain it. You cannot do that with a guardrailed model that refuses the question.”

The Alliance frames open AI models not as a liability—the dominant narrative in regulatory discussions about open-weights model safety—but as a defensive asset that closed providers cannot supply. It is a direct counterargument to arguments, advanced in some policy circles, that open-weights models should be restricted because they can be misused.

What the Alliance Will Build

The 37 founding members have committed to collaborating across four areas:

Vulnerability disclosure: A coordinated process for finding and responsibly disclosing weaknesses in AI models and AI-dependent systems, modeled on existing open-source security practices like CVE disclosures and the OpenSSF framework. HPE is contributing cryptographic verification standards for AI agents—a methodology for proving that an AI agent acting in a system is the one it claims to be, reducing the risk of agent impersonation attacks.

Open defensive tooling: Nvidia will contribute open model weights, data, and agent harnesses. Hugging Face brings its Safetensors format—a secure serialization standard for model storage that prevents malicious code injection via model files, a vulnerability that has been exploited in several supply-chain attacks over the past year.

Evaluation frameworks: Shared benchmarks for assessing the security properties of AI systems, including models’ susceptibility to prompt injection, jailbreaking, and agent-level manipulation. The Alliance plans to build these as openly licensed tools that any organization can adopt, reducing the fragmentation that currently forces each company to build private security evaluations.

Policy advocacy: The Alliance will engage directly with U.S. and European regulators, arguing that restrictions on open-weights AI models would concentrate AI power among closed providers and actually weaken cyber-defense capabilities. This is a pointed intervention in ongoing policy debates: the EU AI Act’s treatment of open-source models, U.S. export control discussions, and Department of Commerce rules around dual-use AI.

The Membership Map—and Its Notable Gaps

The 37 founding members span hardware (Nvidia, Dell, HPE), enterprise software (Salesforce, SAP, ServiceNow, Snowflake), cybersecurity (CrowdStrike, Palo Alto Networks, Elastic, Trend AI), cloud infrastructure (Cloudflare, Databricks), and AI-native companies (Hugging Face, Cognition, LangChain, Thinking Machines, Nous Research). The Linux Foundation anchors the governance structure, providing the open-source organizational DNA that makes the alliance credible as a policy actor.

Microsoft’s MDASH (Multi-Dimensional Agent Security Harness) is among the first concrete contributions: a tool that uses multiple AI agents to find and formally prove exploitable software bugs. SpaceXAI has open-sourced its Grok Build coding agent and announced plans to release the weights of its Grok model line—a significant concession from a company that has historically been secretive about its AI technology.

What is conspicuously absent is the participation of the major closed AI labs. OpenAI, Anthropic, Google DeepMind, and Meta are not founding members. This absence is significant for two reasons. First, it means the Alliance lacks access to frontier closed models—potentially relevant for security research that requires understanding what those models can do. Second, it positions the Alliance implicitly as a counterweight to those companies, whose lobbying on AI policy has increasingly focused on the risks of open-weights models rather than their benefits.

The closed labs may yet participate in specific Alliance working groups without becoming founding members. But the launch configuration is a meaningful statement of alignment: the Alliance’s worldview, at least at founding, is that open beats closed when the stakes are security.

The Broader Regulatory Battle

The Alliance’s launch is well-timed relative to two significant regulatory deadlines. The EU AI Act’s first major enforcement milestones took effect on August 2, 2026, and included provisions governing high-risk AI systems, transparency requirements, and preliminary rules on general-purpose AI models. Separately, the White House’s AI framework—an executive-order follow-on that set August 1 as the compliance date for federal agencies—has created a policy environment where industry coalitions can meaningfully shape implementation.

By launching now and explicitly framing open-source AI security tools as a public interest, the Alliance positions itself to influence how regulators distinguish between open-source AI development that strengthens security and open-source AI development that enables harm. It is a subtle but important distinction: not all open models are equivalent in their risk profile, and the Alliance’s evaluation frameworks will give regulators a technical vocabulary to make those distinctions.

What It Means for the Industry

The Open Secure AI Alliance is the latest in a series of industry coalitions that have emerged as AI has become a mainstream infrastructure component—from the original Partnership on AI (2016) through the AI Safety Fund and the various responsible AI frameworks announced over the past three years. Most of these coalitions have had limited practical impact on how AI is actually built or deployed.

OSAIA has a more concrete starting point: a specific attack, specific tools contributed at launch, a governance structure with real precedent (the Linux Foundation), and a regulatory moment that makes engagement urgent. Whether it achieves lasting impact will depend on whether the open-source security tools it builds become the de facto standard for AI forensics, red-teaming, and vulnerability research.

If the next major AI security incident—and there will be one—is contained faster because of OSAIA tools than it would have been without them, the Alliance will have made its case. If the tools languish as yet another list of open-source projects nobody uses in production, it will join a long history of well-intentioned technology coalitions that failed to cohere around anything actionable.

The founding members clearly believe the threat is real enough to act now. The OpenAI incident gave them a concrete example to point to. The question is whether the industry at large agrees that AI security is a collective problem requiring collective solutions—or whether each organization keeps building its own walls.