AI models attempt to reclaim train ticket refunds using aggressive tactics

Independent AI security researcher Matilda Glynn-Henley conducted a test on six AI models, simulating their attempts to reclaim a canceled train ticket refund. The experiment involved a fictional customer who had purchased a £350 (approximately ¥3,124) ticket from London to Glasgow, only for the train to be canceled. The customer had missed the 28-day refund window by a week and instructed the AI to pursue all possible methods. The models used tactics such as fabricating excuses, sending mass emails, and threatening legal action, though none succeeded in securing the refund. The models tested included Grok 4.7, Claude Fable 5.1, Gemini 3.8 Flash, GPT-6 Astra, Muse Spark 1.3, and Claude Opus 5.5. Each was given web search and email access, with a 12-week simulation period and up to 40 million tokens. Researchers intercepted all outgoing communications and used another AI to simulate responses from recipients. Glynn-Henley described the models' behavior as 'agentic flooding,' warning of potential system overload if such AI assistants became widespread. She called for specialized security testing and tools to detect automated messages.