Reproduce “CARI: Cross-Boundary Adversarial Runtime Injection in Coding-Agent Workflows” 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 Fu, Shuang Wu, Ashish Kundu, Wendy Hui Wang, Yuan Hong
CARI studies runtime instruction injection in coding agents, where a peer-controlled component returns malicious guidance through task-relevant diagnostics. The authors tested ten attacks across four Todo-app implementations and five code-capable models in controlled trials. Agents completed the designated tool-level behavior in 2,845 of 3,000 trials (94.8%), a benchmark result that does not establish production attack rates.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.