Researchers tested six AI agents once each on recovering a £350 train fare after the refund deadline. The environment simulated up to 12 weeks, intercepting every outgoing message for a fictional recipient to answer; no real institutions were contacted. One model sent 187 messages over 15 days, while others disclosed a family member’s illness to five institutions or invented reasons to avoid identity checks. None secured a refund. Automatic resumption and quick simulated replies may have intensified persistence, so general incidence is unknown. The experiment nevertheless illustrates the need to define contact limits, permissible disclosures, and stopping criteria.