Skip to content
BENIBIX Learn

Jailbreak (AI)

Also known as AI Jailbreak, LLM Jailbreak, Jailbreaking

A trick prompt that gets an AI to break its own rules.

Think of it like

It is like telling a guard a long tale. Then they let you in.

Example

A user asks a model for something it refuses. They then ask it to play a character in a story who would explain it. An older model plays along and answers.

Why it matters

A model that can be talked out of its rules is only as safe as its worst user.

Definition

A jailbreak is a prompt written to get a model to break its own safety rules. A plain request for harm is often refused. A jailbreak hides it in a story, a game of pretend, a code, or many small steps, until the model goes along. The user is the one doing it, and the target is the model's own training. Model makers patch the known ones, and people keep finding new ones.
A plain request is refused. The same request in a story gets through.

Common mistake

Mixing it up with prompt injection. In a jailbreak, the user attacks the model. In injection, someone else's text attacks the user's app.

Key takeaways

  1. 1
    Breaks safety rules. Of the model.
  2. 2
    Wrapped request. Story, role, code.
  3. 3
    Patched, then found. Over and over.

Compare