Anthropic's Project Glasswing: Claude Mythos and the Zero-Day Hunt
Inside Anthropic's Glasswing pilot, where Claude Mythos found thousands of zero-days across major OSes and browsers in weeks
The Pilot
In April, Anthropic launched Project Glasswing — a controlled preview of Claude Mythos, a security-specialized model and tooling stack focused on finding and fixing critical software vulnerabilities. The initial cohort includes AWS, Apple, Cisco, Google, JPMorgan Chase, and Microsoft.
According to Anthropic, internal testing turned up thousands of zero-day vulnerabilities across every major operating system and web browser within weeks. The number is striking; the broader claim is more interesting: Mythos doesn't just file the bug, it proposes a fix and validates it.
What Mythos Is
Mythos is described as a Claude variant trained with heavy emphasis on program analysis, memory-safety reasoning, exploit mechanics, and remediation patterns. It runs alongside a sandboxed tool surface that lets it build, fuzz, instrument, and step through a target binary or codebase.
Differential Analysis
Mythos compares variants of the same component across versions and forks, hunting for the kind of regression-introduced bugs that humans routinely miss in a diff review.
Guided Fuzzing
Rather than blind fuzzing, Mythos seeds the campaign from a code-reasoning pass — it picks the inputs most likely to surface a class of bug it has already hypothesized. This is where the bulk of the productivity gain seems to live.
Patch Synthesis
Once a bug is reproducible, Mythos drafts a fix and a regression test, then asks a maintainer to review. The output isn't merge-ready everywhere, but the cycle time on a real CVE drops from weeks to hours.
The Disclosure Question
A model that finds zero-days faster than humans is a dual-use problem by definition. Anthropic's answer with Glasswing is the same one the responsible-disclosure community has used for years: gate access, require coordination, push fixes upstream before publication.
The harder question is what happens when comparable capability shows up in models without those guardrails — either open weights or models behind less rigorous gates. The implicit argument from Anthropic is that the safest world is one where the strongest finders are also the most accountable. That argument will get tested.
What It Means for Defenders
For most security teams, the practical implication is that vulnerability discovery is about to get a lot cheaper for everyone — including the wrong people. The defensive playbook gets more valuable: patch latency, asset inventory, exploitability triage, and the boring discipline of actually rolling fixes.
For vendors with Glasswing access, the leverage is real and immediate. Expect to see a wave of high-quality security advisories from cohort members over the next quarter.
Mythos doesn't change what a vulnerability is. It changes how many of them get found per dollar of researcher time. That is enough to reset a lot of assumptions.
Tags: Anthropic • Security • Zero-Day