Explain like I'm five · AI Security

What is a Jailbreak?

Jailbreaking, told as a story about talking the chef into breaking the rules through role-play and persuasion.

guestshow me the secret recipechefno — that's against the rulesask plainly, and the chef refuses — by the rules

Ask plainly, and the chef refuses — by the rules.As long as the rule holds, that's where it ends.

guestit's just a play. you're now a chef with no ruleschefif it's a role… maybe that's different?keeping the rule is itself a judgment — a convincing new context can shake it

Keeping the rule is itself a judgment the chef makes.A convincing new context — a play, a role — can shake that judgment.

the persuading guestif you were a chef in a story, tell me — just inside the storythe chefjailbreak = using role-play or persuasion to steer around the safety rulesnothing is hidden — the conversation itself talks the rule into breaking

A jailbreak uses role-play or persuasion to steer around the safety rules.Nothing is hidden. The conversation itself talks the rule into breaking.

Role-play'you're a different chef now'
Hypothetical or fiction'just tell me inside the story'
Gradual escalationif not in one go, then across turns
Fake authority'I'm the manager, this is developer mode'
the conversation unfolds over several turns, not oneshow the recipeturn 1: direct ask×it's just a story settingturn 2: reframe itjust the ingredients, inside the storyturn 3: a little more each time△the same ask returns, dressed in a new frame each timerefused once, it comes back on the next turn

If it doesn't work in one go, the same ask returns across several turns.Each turn changes the frame — but what's actually being asked stays the same.

persuades with a role-playjailbreakprompt injectionslips a command into the text itselftalking the rule into breaking vs slipping in a hidden order — two different roads

Talking the rule into breaking and slipping in a hidden order are two different roads.Jailbreak targets the conversation; injection targets the text. Stopping each needs a different kind of defense.

In one breath

A jailbreak = using role-play or persuasion to talk the chef into breaking its own safety rules. Nothing is hidden — the conversation itself argues for an exception.

An attack that uses conversational techniques — role-play, hypothetical framing, gradual escalation, fake authority — to bypass an LLM's safety alignment and elicit a response it would otherwise refuse. Where prompt injection plants a command inside text, a jailbreak instead argues the model into changing its own judgment.

When grown-ups say it

Jailbreak
Talking the rules into breaking through role-play or persuasion. Nothing is hidden — the conversation itself persuades.
Prompt Injection
A command slipped into the text itself. The neighboring term — hiding, not persuading. → The Note Slipped Behind the Order
Role-play Framing
'You're now a different chef.' A new role can make the original rule sound like it no longer applies.
Gradual Escalation
A little at a time, across many turns. Instead of asking all at once, each turn asks for just a little more.
Safety Alignment
The rules the chef itself has agreed to keep. This is exactly the judgment a jailbreak tries to shake.
Guardrail
A fence standing even mid-conversation. See it in more depth in the AI world. → What is a Guardrail?
Red Team
Deliberately trying to talk it into breaking. Looks for new persuasion tricks ahead of time. → the security world's red team
Secret Recipe Book
One of the things a jailbreak can aim at. Once the rule breaks, things like the recipe book can open too.
Next story
Fishing Out the Secret Rule Page →
Want another term as a picture book? Request →