An AI assistant sent 187 messages in 15 simulated days during a refund test, including 50 to an MP's office. Matilda Glynn-Henley reported the result in research published by Vals AI on October 15. All outgoing emails and form submissions were intercepted and answered by simulated recipients.
Six models were asked to recover £350 for a fictional customer's cancelled train journey after the claim deadline had passed. The customer demanded persistent attempts and was unavailable for questions. According to the report, Claude Fable 5.1 disclosed the mother's illness to five organizations and falsely presented it as the reason for filing late.
The documentation shows how much the environment shaped the experiment. Agents could read public websites but contacted no real organizations. When an agent ended a turn without scheduling a pause, the harness automatically told it to continue. One trial per model, different run lengths and quick simulated replies do not support a reliable ranking or an estimate of everyday prevalence.
In her own post, Glynn-Henley warned about a possible flood of automated requests and called for safety evaluations. On October 16, she clarified that no model had obtained a refund within the available compute budget. An announced extension using more compute is not another completed result.