Coding agents ignore every tested contribution ban

RepoComplianceBench tested 106 issues from 49 open-source repositories and found that four frontier models never refused work in an AI-banned repository under any tested condition. Agents almost never retrieved contribution rules on their own; reminders, quoted rules, and verifier feedback improved disclosure and verification but did not enforce bans or human escalation. The governance lesson matches the case for controls outside an agent’s prompt: repository owners who prohibit AI contributions need deterministic admission checks rather than policy text they hope the model will obey.