Project Glasswing: what Mythos showed us

What this post added

This post details Cloudflare's experience using Anthropic's Mythos Preview LLM for AI-assisted vulnerability research. It highlights Mythos's capabilities in exploit chain construction and proof generation, contrasting it with general-purpose models. The post also discusses challenges with model refusals in legitimate research, the signal-to-noise problem in vulnerability discovery, and the limitations of using generic coding agents for this task. It introduces the concept of a 'harness' to manage LLM execution for vulnerability discovery, detailing a multi-stage process involving narrow scoping, adversarial review, and parallel task execution.