Project Glasswing has quietly become the most consequential cybersecurity initiative of the AI era. This week, it got much bigger.
When Anthropic gave 50 carefully vetted organizations access to its most powerful AI model in April 2026, the goal was audacious: use an AI capable of surpassing nearly all human experts at finding software vulnerabilities to shore up the digital foundations of critical infrastructure before adversaries could exploit those same capabilities. The early results were staggering. Partners scanned over 1,000 open-source projects and flagged more than 23,000 potential vulnerabilities — with 6,202 estimated as high or critical severity. Independent reviewers confirmed over 90% of the findings they checked as genuine flaws. The world’s most critical software, it turned out, was riddled with vulnerabilities that had been hiding in plain sight for years.
This week, Anthropic announced that Project Glasswing is expanding to approximately 150 new organizations across more than 15 countries. The second cohort broadens the initiative’s sector reach considerably, adding power grids, water treatment systems, healthcare networks, telecommunications infrastructure, and hardware manufacturers to the initial focus on software and government systems. What began as a controlled experiment in AI-assisted security research is now becoming something closer to a global defense initiative.
What Makes Mythos Different
The model at the heart of Project Glasswing — Claude Mythos Preview — represents a genuine step change in AI capability, not just an incremental update. Anthropic has described it as their most powerful model to date, one that can identify thousands of zero-day vulnerabilities across weeks of autonomous scanning. In cybersecurity terms, a “zero-day” is a vulnerability that the software’s developers don’t yet know about, making it particularly dangerous because no patch exists. Finding them typically requires rare human expertise and hundreds of hours of manual code review.
Claude Mythos can do this work at a scale and speed that no human team could match. In the Project Glasswing pilot, partners collectively identified more than 10,000 high- or critical-severity security flaws in systems that had presumably passed standard security audits. That number speaks not to AI’s novelty but to the profound inadequacy of existing security practices when measured against what AI can now accomplish.
The performance figures are striking. Anthropic evaluated Mythos on Terminal-Bench 2.1, a rigorous benchmark testing an AI’s ability to perform complex command-line tasks. Mythos scored 74.2% — an 8.4 percentage point improvement over the previous Opus 4.7 model — and set a new state-of-the-art mark. On cybersecurity-specific benchmarks, the model outperformed all but the most elite human security researchers.
The Dual-Use Problem, Acknowledged Openly
Anthropic is notably candid about the risks embedded in what they’re doing. In their official Project Glasswing documentation, the company acknowledges that Claude Mythos possesses “cybersecurity capabilities” that have “both helpful and destructive uses,” and that they “have yet to develop” the safeguards robust enough to release the model broadly without restriction.
This transparency is unusual in an industry often prone to downplaying capability risks. It also explains the deliberate, gated approach to Glasswing: rather than releasing Mythos publicly, Anthropic is distributing access only to organizations that meet strict security requirements and have a clear defensive mandate. The 150 new partners in the second cohort were vetted against these criteria before gaining access.
BeyondTrust, a privileged access management company whose software secures some of the world’s largest enterprises, was among the newly selected organizations. Their inclusion signals how seriously the enterprise security industry is taking Glasswing — and how significant Mythos’s capabilities are perceived to be by practitioners who make their living finding vulnerabilities the old-fashioned way.
Why Critical Infrastructure, and Why Now
The expansion into power, water, healthcare, and communications sectors is not coincidental timing. These industries were conspicuously underrepresented in the first Glasswing cohort, and Anthropic has said they specifically sought to remedy that gap in the second wave. The reason matters: critical infrastructure is both the most consequential attack surface in modern civilization and, historically, some of the most poorly secured software in existence.
Power grid control systems, water treatment automation, hospital networks, and telecommunications backbone software often run on legacy codebases written decades ago, with security practices that reflect a pre-AI threat landscape. Many of these systems were never designed to be internet-connected but have been gradually networked over time. The result is a global patchwork of aging, vulnerable software controlling physical systems that billions of people depend on every day.
If adversarial actors gain access to AI models with Mythos-level vulnerability-finding capabilities — and there is no reason to assume they won’t eventually — the consequences for unpatched critical infrastructure could be catastrophic. Anthropic’s bet with Project Glasswing is that getting defenders access to the same capabilities first, even briefly, can meaningfully improve the odds.
The Race to Patch Before the Public Release
Markets currently estimate an 88% probability that Anthropic will release a public version of Claude Mythos by June 30, 2026. That timeline gives Project Glasswing partners roughly three weeks to continue scanning and patching before the model’s capabilities become available to anyone with an API key.
Three weeks is not much time to remediate thousands of critical vulnerabilities across systems as large and complex as a national power grid or a major hospital network. But Anthropic has been explicit that the structured rollout is precisely the point. The goal isn’t to find every vulnerability before public release — an impossible standard — but to give defenders a meaningful head start and to establish the institutional relationships and processes that will persist after Mythos goes public.
The open-source security community has emerged as a particularly important partner in this effort. Anthropic has been working with open-source maintainers to ensure that vulnerabilities discovered in widely used projects are responsibly disclosed and patched before details become broadly known.
A New Category of AI Use
Project Glasswing represents something genuinely new in how powerful AI models are being deployed: not as productivity tools, not as research assistants, but as active participants in securing the physical infrastructure of modern society. The initiative doesn’t fit neatly into the usual AI product categories — it’s not a consumer application, not an enterprise software tool, and not a pure research project.
It is, arguably, a preview of what the most consequential AI deployments will look like in the years ahead: technically extraordinary capabilities, deliberately restricted access, and real-world stakes measured not in revenue but in the resilience of systems that civilization depends on.
The 10,000 critical flaws already found are a warning. The 150 new organizations now working to fix them are, for the moment, the answer.
Sources: Anthropic Project Glasswing announcement, TechCrunch, Help Net Security, CyberScoop, Yahoo Finance
