Launder-Bench benchmark for detecting agentic code laundering
Original title: Launder-Bench: Can Repository-Level Agentic Code Laundering Be Detected?
Introduces Launder-Bench with 1,128 pairwise and 229 attribution instances to test provenance under agentic rewriting from 282 repositories.
The paper defines repository-level agentic code laundering and presents Launder-Bench, the first such benchmark. Launder-Bench-Pair contains 1,128 instances for provenance verification and Launder-Bench-Attr contains 229 source-attribution queries. A source-hidden reconstruction protocol with execution verification is used; existing static and embedding methods fall short of saturating the benchmark.
- Benchmark constructed from 282 real-world repositories
- Launder-Bench-Pair: 1,128 pairwise provenance queries
- Launder-Bench-Attr: 229 source-attribution instances
- Strongest static method reaches 0.66 AUROC
- Repository encoders rank true source first on up to 22% of queries
From the original article
1 min read at arxiv.orgCoding agents can now rewrite repository-scale codebases while preserving tested behavior, creating a new challenge for plagiarism detection, license auditing, and software provenance. We ask whether such rewrites can still be traced to their source. Existing clone benchmarks emphasize fragments and bounded edits, while repository-transformation benchmarks evaluate functional correctness rather than provenance.
By Yuntong Hu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.