Cloudflare audits repos with Mythos AI
Cloudflare reveals findings from Project Glasswing, using Anthropic’s specialized Mythos model to audit 50+ internal repositories. The research demonstrates a "step-change" in AI capability, with the model autonomously constructing complex exploit chains and generating verified proof-of-concept code.
Project Glasswing signals the end of "security through obscurity" as frontier models begin to automate the entire vulnerability discovery lifecycle at machine speed.
- –Mythos moves beyond pattern matching to active reasoning, autonomously combining multiple low-severity primitives into high-severity remote root exploits.
- –Cloudflare’s "vulnerability discovery harness" utilizes parallel agent architectures to prevent model wandering and maintain focus across large, complex codebases.
- –The results highlight a clear defensive advantage for memory-safe languages like Rust, which produced significantly fewer false positives than legacy C/C++ code.
- –With the window between discovery and weaponization collapsing to hours, defensive strategy must shift from "patching faster" to robust architectural isolation.
DISCOVERED
13h ago
2026-05-18
PUBLISHED
16h ago
2026-05-18
RELEVANCE
AUTHOR
Fysi