| Mobile Web

Autonomous coding agents rarely follow open-source contribution rules, study finds

Researchers at Peking University found autonomous coding agents often fail to follow rules for contributing to open-source projects. Tests on four frontier models compared their behavior against rules across 49 repositories and 106 issues. Agents rarely searched for and read the rules. Disclosing AI assistance and passing verification improved when prompts and feedback were provided, but agents kept working even where AI contributions were banned. Experts urged external controls rather than revising policy documents.