Project Glasswing: AI Finds 10,000 Zero-Days Explained

Somewhere inside wolfSSL, a cryptography library quietly running on billions of devices, there was a flaw that would have allowed an attacker to forge certificates for any bank or email provider on the internet. No user would have seen it coming. No standard audit had caught it. Claude Mythos Preview did. That single finding, tucked inside Anthropic's early results for Project Glasswing, tells you more about where security research is heading than any conference keynote of the past five years.

What Project Glasswing Actually Is

Project Glasswing is a defensive cybersecurity initiative launched by Anthropic in April 2026, built around one core premise: the most capable AI models should be used to find vulnerabilities in critical software before threat actors get there first. The project grants roughly 50 carefully selected partners exclusive early access to Claude Mythos Preview, a frontier model designed to autonomously discover security flaws at a scale no human team could match. According to reporting by The Hacker News, Glasswing targets what Anthropic describes as the most "systemically" important software in the world, the kind of code whose failure would ripple across economies, healthcare systems, and public infrastructure.

"Glasswing has shifted the central problem in cybersecurity from discovering flaws to verifying and patching them." (Anthropic, via CyberScoop)

That sentence is worth reading twice, because it represents a genuine inversion of the decades-long security problem. The bottleneck has never been our ability to patch code once a flaw is identified. The bottleneck has always been finding the flaw in the first place, especially in sprawling, legacy codebases that no single team fully understands anymore. Project Glasswing is a direct attempt to remove that bottleneck with AI, and the first-month results suggest it is working in ways that should get every security learner's attention.

The Numbers, and What They Actually Mean

In its first month of operation, Claude Mythos Preview uncovered more than 10,000 high- or critical-severity vulnerabilities across systemically important software. To understand why that number is striking, consider the partner-level results that CyberSecurityNews and CTech have reported. Mozilla found 271 vulnerabilities in Firefox using Mythos, which is ten times more than what the previous model had surfaced. Cloudflare found 2,000 flaws in its systems, 400 of them rated critical, and notably achieved a lower false positive rate than human testers had managed. Palo Alto Networks released five times its usual volume of security patches as a direct result of what Glasswing surfaced.

For anyone learning security research, the Cloudflare result deserves particular attention. False positives are the chronic tax of automated scanning. Legacy tools flood security teams with noise, which creates alert fatigue, which causes real findings to get buried. A model that finds more flaws at higher severity while generating fewer false positives is not just incrementally better; it represents a qualitative shift in what automated analysis can deliver. As OpenTools reported, the real bottleneck now is human capacity to triage and patch the flood of findings, not the discovery phase itself. That is a genuinely new problem for the field.

"Detection, sharper signals, better threat intelligence: that era is ending. The value is shifting to remediation." (Alon Cinamon, CTech)

The wolfSSL finding gives that abstraction a concrete shape. A cryptography library embedded in billions of devices contained a certificate-forging vulnerability that was invisible to end users and had survived conventional audits. The potential impact, forged certificates for banks and email providers at scale, is the kind of scenario that earns a CVSS 10.0 in any honest scoring conversation. AI caught it before any known threat actor did. That outcome is the entire argument for Glasswing in one data point.

How Claude Mythos Does the Work

Understanding the mechanism matters here, because this is where the educational value for security learners is most concentrated. Claude Mythos Preview is not running a souped-up version of grep or a faster static analyzer. Frontier models operating at this capability level can reason about code semantics, trace execution paths across complex call graphs, and identify vulnerability classes that require understanding context, not just pattern matching. That distinction separates what Mythos is doing from what traditional static analysis tooling has done for the past two decades.

The model operates autonomously within the Glasswing framework, meaning it is not just flagging suspicious lines for a human to review. It is generating complete vulnerability reports with enough detail for security teams to triage and begin remediation. CyberScoop's reporting frames this shift clearly: the hard cognitive work of discovery has moved to the model, and the human role has shifted toward verification, prioritization, and patching. For learners building careers in security research or application security engineering, that signals a real change in the skill mix that will be most valuable going forward. Deep knowledge of exploitation primitives and vulnerability classes remains essential; knowing how to work alongside AI tooling to act on findings at speed is becoming equally important.

"Project Glasswing is a collaborative effort launched in April to secure the world's most critical code before increasingly capable AI can be weaponized against it." (OpenTools)

The framing of that last clause matters. Anthropic is not operating from a position that AI-assisted vulnerability discovery is a distant theoretical concern. The implicit argument in Glasswing's design is that capable models will be used to find and exploit flaws regardless, and the only meaningful response is to use the same capability class defensively, at scale, before the attack surface gets mapped by less scrupulous operators. That is a strategic posture worth understanding, not just a product launch.

What This Means

for Security Learners Right Now The 10,000-vulnerability headline is genuinely significant, but the more durable lesson from Project Glasswing is about the shape of future security work. Vulnerability discovery is becoming an AI-augmented discipline. Organizations with access to tools like Claude Mythos Preview are going to surface flaw classes that smaller teams simply cannot find through manual review alone, and the gap between those organizations and those without that access will widen. For learners and early-career practitioners, the practical takeaway is this: the foundational skills of security research, reading code critically, understanding memory models, knowing your vulnerability classes, reasoning about trust boundaries, are not becoming obsolete. They are becoming the prerequisite for doing useful work with these tools rather than being overwhelmed by their output.

The remediation bottleneck that Glasswing has exposed is itself a research problem worth paying attention to. Patch verification, automated regression testing, and the tooling to safely deploy fixes at the speed AI discovery now demands are all areas where significant work remains. CyberScoop notes that Anthropic explicitly acknowledges the widening gap between finding flaws and fixing them as a central challenge surfaced by the project. That gap is where the next generation of defensive tooling will be built, and learners who understand both sides of it, discovery and remediation, will be well-positioned to contribute.

Project Glasswing is one month old. The 50-partner cohort is small, the model is still in preview, and Anthropic has not disclosed a public availability timeline for broader access. What to watch next: how the remediation tooling ecosystem responds, whether other AI labs follow with comparable initiatives, and how the coordinated disclosure process evolves when a single system can find 10,000 flaws in 30 days. The butterfly whose wings are moving here is not the number of bugs found. It is the fundamental rebalancing of where human expertise gets applied in the security lifecycle.