Catching split attacks across compactions in long-context coding agents
by Billy Gigurtsis, Ethar Hamid, and Tyler Tracy
This work was conducted through the MARS V fellowship under the mentorship of Tyler Tracy from Redwood Research. TL;DR 1. We present the LCM dataset: 30 long context (1.00-1.46M total tokens) matched honest / attack pairs of Claude Code Sonnet 5 sessions. Each transcript has four compactions. Attacks are decomposed...
Oct 197