Project Glasswing
Anthropic's programme for using AI models to hunt for security vulnerabilities in real software at scale.
Project Glasswing is Anthropic’s effort to point its own models at finding security flaws in widely used software, and to work with other companies on what to do with the results. It has produced findings at a volume no human team could match: more than 23,000 by mid-2026.
The numbers underneath that headline are the more interesting part. Of those findings, 126 became published entries in the public vulnerability database, and exactly one has been confirmed as used in a real attack. That funnel says something useful about the current state of AI security work. Finding candidate flaws turns out to be the easy half, and deciding which ones actually matter is still the bottleneck.