Reproduce “Stopping Is Not the Limit: Understanding and Eliciting Agent Capability in Long-Horizon Vulnerability Exploitation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Deokhyung Kang, Minjae Gwon, Sangdon Park

SCoRe extends vulnerability-exploitation agents beyond native stopping by restarting exploration when stops recur and marking unachieved capabilities judged impossible as unproven. On the 21-task Mythos-ACE subset, MiMo-V2.6-Pro reaches Cov@3 coverage of 67.0% with SCoRe, versus 53.3% with summary-based continuation and 32.4% with native stopping. These results are under a shared 2,000-turn ceiling and establish higher verified coverage within this tested setting.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.