Wintermute rebuilds Molly agent after failure

The author details ongoing developments in building Molly, an offensive-security agent designed with a Rust-and-Python stack where the language model reasons about hypotheses while deterministic code executes the actions. After encountering poor results and slow execution times on an XBOW vulnerability benchmark, the author implemented a skeptic agent named deck to enforce rigorous measurement and guardrails. This setup ultimately enabled Molly to successfully solve a complex HTTP request-smuggling challenge involving multiple proxy version discrepancies. Open article →