Jailbreak

What is an AI jailbreak?

"Jailbreak" used to mean hacking a phone to escape the manufacturer's restrictions. In AI, it means crafting prompts that slip past a model's safety guardrails so it says things it's supposed to refuse — how to make something dangerous, illegal advice, or content it's been told to keep quiet about.

How is it different from a hack?

A jailbreak usually needs no code exploit and never touches a server. You do it purely by talking. Attackers aren't hunting for software bugs — they're hunting for bugs in the model's "understanding," using roleplay, rephrasing and multi-step coaxing to convince the model it's in a scenario where it's allowed to speak freely.

Common jailbreak moves

Roleplay
Tell the AI to act as "an AI with no rules," then extract the harmful stuff.
Storytelling
Wrap harmful content as "fiction" or a "case study" to dodge sensitive words.
Incremental nudging
Start with harmless questions and steer the conversation toward the forbidden.
Obfuscation
Hide intent with ciphers, homophones or translations.

Why is it so hard to stop?

At its core, jailbreaking exploits how flexible language is. A model can never enumerate every dangerous phrasing — patch one and attackers find another. It's a permanent cat-and-mouse game: safety teams keep patching, jailbreakers keep finding new routes.

Bottom line: an AI jailbreak doesn't break into the system — it talks past the guardrails until they don't matter.