Reproduce “PyCache Trap: The Inspection-Execution Gap in Agent Skill Scanners” 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
Jie Liao, Simeng Qin, Wenqi Ren, Wei Zhou, Junhao Wen, Ranjie Duan, Yang Liu, Xiaojun Jia
PyCache Trap tests whether Python executes a bundled cache that differs from a Skill’s visible source, while scanners inspect only readable files. EAV links runtime-reachable artifacts to inspected source and compares caches with trusted, interpreter-matched rebuilds. In the Python 3.11 setup, EAV detected all 100 source-present substitutions; its benchmark reached 92.8% recall and 10.0% false-positive rate, not universal coverage.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.