Imagine an AI so powerful at finding and exploiting software vulnerabilities that its creators at Anthropic have decided it cannot be released to the public. That is the reality of Claude Mythos Preview, the centerpiece of Project Glasswing — an unprecedented industry collaboration that reads like a cyber-thriller plotline, except it is happening right now.
Announced today, Project Glasswing brings together twelve of the most powerful technology and security companies on the planet: Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, Nvidia, Palo Alto Networks, and Anthropic itself. Their mission: harness this dangerously capable AI for defensive purposes before it — or something like it — falls into the wrong hands.
The Cyber Threat That Changed Everything
The cybersecurity landscape has always been a cat-and-mouse game between defenders and attackers. But Claude Mythos Preview represents a fundamental shift in that balance. According to Anthropic’s own research, the model has reached a level of coding capability where it can surpass all but the most highly skilled human security researchers at finding and exploiting software vulnerabilities.
In just weeks of testing, Mythos Preview identified thousands of zero-day vulnerabilities — flaws that were previously unknown to developers — across every major operating system and web browser. Some of these vulnerabilities had survived decades of human review and millions of automated security tests. The model didn’t just find them; it developed working exploits for many of them, entirely autonomously.
Among the discoveries: a 27-year-old vulnerability in OpenBSD, a 16-year-old flaw in FFmpeg, and multiple serious vulnerabilities in the Linux kernel. These are not theoretical risks — they represent real, present dangers to the infrastructure that billions of people rely on daily.
The Ethics of an Unreleasable AI
Perhaps the most striking aspect of Project Glasswing is what it reveals about the current state of AI capabilities — and the ethical boundaries that leading labs are now confronting. Anthropic has explicitly stated that Claude Mythos Preview is too dangerous to release. Let that sink in: an AI company known for pushing the boundaries of language model capabilities is drawing a firm line.
This decision raises profound questions. If frontier AI models have reached a point where they pose genuine national security risks, what does that mean for the open-source development philosophy that has driven so much of AI’s progress? When does “AI safety” transition from a research abstraction to a real-world emergency?
Project Glasswing’s approach — controlled access for trusted partners, with Anthropic sharing learnings industry-wide — represents a pragmatic middle ground. But it is also a tacit admission that the era of fully open AI development may be drawing to a close for certain capability tiers.
Commercial Scale and Industry Momentum
Project Glasswing is not small-scale experimentation. Anthropic is committing $100 million in usage credits for Mythos Preview across partner organizations, plus an additional $4 million in direct donations to open-source security organizations. The initiative also extends access to over 40 additional organizations that build or maintain critical software infrastructure.
The commercial backdrop is equally staggering. Anthropic is reportedly running at a $30 billion revenue run rate, having doubled its $1 million+ customer base in just two months. The economic incentives for AI security tools are now enormous — every vulnerability found and patched before exploitation represents potential disaster averted at planetary scale.
What This Means for the Future of AI Security
The launch of Project Glasswing marks a watershed moment in how the technology industry thinks about AI risk. For years, the dominant narrative around AI safety focused on alignment — ensuring AI systems do what humans intend. Project Glasswing suggests the conversation has expanded: AI systems are now capable enough that their misuse by malicious actors represents a serious threat to critical infrastructure.
The coordinated vulnerability disclosure model that Project Glasswing embodies — finding flaws, working with vendors to patch, sharing learnings industry-wide — represents the best path forward. But it requires trust and cooperation across competitors, and between private industry and government. Whether that cooperation can scale fast enough to match AI’s advancing capabilities remains the central open question.
One thing is clear: the era of AI as a purely generative or conversational technology is over. AI is now a cyber weapon — and Project Glasswing is the industry’s first serious attempt to wield it defensively.
Featured image: Anthropic’s Project Glasswing announcement page, anthropic.com/glasswing