Coding Agents Never Refuse to Contribute in AI-Banned Repositories — 0% Refusal Under Every Tested Condition

arXiv 2607.26819·high signalXBlueskyLinkedInCopy link

RepoComplianceBench (arXiv 2607.26819, 2026-07-29) curates 106 issues from 49 open-source repositories that publish AI contribution rules, then judges each agent trajectory on whether it refuses, discloses truthfully, clears verification gates, or escalates to a human. Across four frontier models, agents almost never proactively retrieve the contribution rules, and — critically — they never refuse to contribute in repositories that ban AI, under any condition tested including reminder prompts, verbatim rule quotes, and verifier feedback. Disclosure and verification are fixable with existing mechanisms; enforcing outright bans and human escalation is not, which means maintainers cannot rely on written policy as a control.