Launder-Bench: Can Repository-Level Agentic Code Laundering Be Detected?
| Source: arXiv AI
Tags: Launder-Bench, code provenance, coding agents, license compliance, plagiarism detection
Launder-Bench tests whether code rewritten by coding agents can be traced to its source repository; the best static detector reaches only 0.66 AUROC, exposing a gap for license and provenance audits.
Details
Coding agents can now rewrite whole repositories while preserving tested behavior, which complicates plagiarism detection, license auditing and provenance. The authors define repository-level agentic code laundering and build Launder-Bench from 282 real repositories. A source-hidden reconstruction protocol rebuilds modules and structure and keeps rewrites only after execution-based verification. Launder-Bench-Pair has 1,128 pairwise instances, and Launder-Bench-Attr has 229 source-attribution instances with abstention allowed. The strongest static method hits 0.66 AUROC on pairwise detection. Repository encoders rank the true source first on up to 22% of attribution queries but earn calibrated credit on only 1-2%. For legal and compliance teams, current similarity tools are not enough to prove derivation.