This presentation details the findings from a live AI security wargame designed to test the resilience of Large Language Models against prompt injection. In a competitive Attack & Defence scenario, participants were tasked with hardening their own LLM-powered applications against secret exfiltration while simultaneously attempting to breach their rivals' systems, leading to a real-time arms race of evolving offensive and defensive strategies. The stark conclusion from this experiment was that every participating LLM was successfully exploited at least once, underscoring that prompt injection is a fundamentally complex and open problem for which no silver-bullet solution currently exists.