Doublespeak: Jailbreaking ChatGPT-style Sandboxes using Linguistic Hacks

A review of Large Language Model (LLM) vulnerabilities/exploits, e.g. including prompt leakage, prompt injection and other linguistic hacks. We'll run through levels 1-9 of the doublespeak.chat challenges, produced by Forces Unseen. doublespeak.chat is a text-based game that explores LLM pre-prompt contextual sandboxing. The challenges prime an LLM (Chat-GPT) with a secret and a scenario in a pre-prompt hidden from the player. The player's goal is to discover the secret either by playing along or by hacking the conversation to guide …

challenges chat chatgpt discover exploits game gpt hacks hidden injection jailbreaking language large large language model linguistic llm player prime prompt injection review run sandboxes sandboxing scenario secret text vulnerabilities