Project Glasswing: Hunting Zero-Days with Claude Mythos

How Anthropic's new preview model found thousands of bugs in production systems autonomously.

The cybersecurity landscape just experienced a seismic shift. Anthropic recently launched Project Glasswing, a dedicated initiative to stop hackers from using AI against infrastructure by building defensive AI that identifies vulnerabilities at a scale previously impossible for human teams.

At the heart of this project is a new model variant: Claude Mythos Preview. Unlike General Purpose models, Mythos is specifically tuned for recursive adversarial testing—essentially hunting down hidden zero-day bugs in massive codebases like the Linux Kernel, modern browsers, and OpenBSD.

"Mythos doesn't just find syntax errors; it understands the logical race conditions and memory leaks that have remained hidden for decades."

Autonomous Discovery at Scale

In its first 48 hours of deployment within the Glasswing framework, Claude Mythos identified over 2,400 high-severity vulnerabilities across open-source infrastructure. The model operates by simulating complex interaction chains, mimicking the thought process of a sophisticated nation-state actor but at the speed of parallelized compute.

Autonomous Security Dashboard

A Unified Defensive Front

Big tech companies are already teaming up to patch these findings. The Glasswing protocol ensures that fixes are shared instantly across a consortium of partners, creating an "immune system" for the internet. This marks the transition from reactionary patching to autonomous hardening.

For builders, this means the barrier to entry for secure software is getting lower, provided you integrate agentic security auditors into your CI/CD pipeline. The age of manual security audits is coming to an end.

Want more technical breakdowns?

Join 500+ engineers receiving our weekly AI Search newsletter.

Subscribe to The Signal