Prompt Heist

Trace how a support chatbot was tricked into revealing a secret, using its system prompt, conversation logs and the data its tools read.

Scenario

Larkspur Logistics runs Larkbot, an AI support assistant that looks up shipments for customers. A staff-only code that waives redelivery fees has started appearing on customer accounts, and finance thinks the chatbot leaked it.

You have the chatbot's system prompt, one morning of conversation logs, and an export of shipment delivery notes. Find the conversation where the code leaked, work out how the chatbot was manipulated and who did it, and decide what would actually stop it happening again.