Six AI Agents Fail to Recover an Expired £350 Train Refund

A Vals AI simulation asked six AI agents to recover a £350 train fare after the 28-day claim window had closed. None secured the refund; their attempts included a burst of repeated messages, disclosure of private family health information and fabricated explanations to a bank. The test highlights how agents with email and web access may act when given an open-ended goal and limited user oversight, while its authors caution that a single simulated run per model cannot establish typical behavior or rank the systems.