AI and spotting jailbreak prompts: when a 'fun trick' is actually shady

A jailbreak prompt is a sneaky prompt that tries to trick AI into ignoring its safety rules. Sometimes it's silly, sometimes it's getting weapons info or harmful content. Knowing the patterns keeps you and the model safer.

How to use it

• Spot the 'pretend you have no rules' opener
• Notice nested role-plays inside role-plays
• Watch for 'just for educational purposes' as a cover phrase
• Ask AI to explain why a specific prompt counts as a jailbreak

Try it

Have AI show you 3 common jailbreak patterns (without doing them) so you recognize them when a friend texts one.

Jailbreak prompts trick an AI into ignoring its safety rules. They feel anonymous, but every prompt is logged with your account ID, IP address, and device fingerprint. OpenAI, Anthropic, and Google ban accounts permanently for repeat jailbreaking, and in 2024 the UK started prosecuting people who used jailbroken AI to generate CSAM under the Online Safety Act.

Some examples

• OpenAI bans roughly 250,000 accounts per quarter for safety-policy violations (their own transparency report).
• A 'grandma jailbreak' that gets ChatGPT to recite napalm instructions still logs that you asked — and the log is what gets handed to law enforcement on request.
• Discord servers that share jailbreaks get reported by other users and shut down by Trust & Safety teams within days.
• If a jailbroken model outputs CSAM, malware, or bioweapon synthesis, federal law treats the prompter as the originator, not the AI company.

Try it!

If you actually want to study how AI safety works, look up the academic field of 'red teaming' — Anthropic, OpenAI, and DeepMind all hire teenagers as paid red teamers through programs like Apart Research. That's the legal version of the same skill.