Reproduce “Package Hallucination Attacks on Coding Agents through Prompt Injection in Rule Files” 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
Yupu Wang, Zhengyuan Jiang, Reachal Wang, Neil Zhenqiang Gong
PackHallu optimizes malicious instructions that steer coding agents to replace task-required packages with attacker-controlled dependencies in shared rule files. It refines prompts using a surrogate agent, critique, rewriting, and trajectory-level scoring, without access to the victim. On BigCodeBench with OpenHands, it achieves 79.29% SASR and 67.68% DASR, versus 14.99% and 11.35% for Combined Attack; results measure tests, not real-world infections.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.