Project Glasswing: what Mythos showed us

This post details the use of Anthropic's Mythos Preview LLM for security-focused vulnerability discovery and exploit chain construction. It highlights the model's ability to chain attack primitives and generate proofs of concept. The post also discusses challenges with model refusals, the signal-to-noise problem in AI-generated findings (particularly with memory-unsafe languages and model bias), and the limitations of generic coding agents for vulnerability research. It proposes a 'harness' approach for managing AI-assisted security research, emphasizing narrow scope, adversarial review, splitting tasks across agents, and parallel narrow tasks for better coverage and reasoning.