Splitting a CTF Task Into Isolated Sub-Contexts Lets a Local gemma-4 Solve 18.52% of Challenges Standard Agent Loops Fail

arXiv 2609.12839·low signalXBlueskyLinkedInCopy link

Open-weight small language models deployed locally bypass proprietary API guardrails, but as autonomous agents they collapse on long exploratory work because accumulated tool-call output bloats context. Context segmentation is a two-level framework that divides exploitation tasks into contextually isolated sub-problems, evaluated on picoCTF with memory-constrained gemma-4 models. For the E4B model it behaves like an intelligent search, matching brute-force retry rewards at better token efficiency and solving 18.52% of tasks standard agentic execution cannot complete. Presented at the RAISE 2026 workshop at ESORICS, non-archival, with code released.