AIBound

Home / Glossary / AI jailbreak

Glossary

What is an AI jailbreak?

An AI jailbreak is a prompt, or a series of prompts, that gets an AI model to ignore the safety rules it was built with. The model then produces output or takes steps it would normally refuse. OWASP treats jailbreaking as a form of prompt injection.

By the AIBound team · Updated 17 November 2026

Key takeaways

  • A jailbreak targets the model's own safety rules, and the person typing is usually the one attempting it.
  • Model makers keep updating their safety training, and new jailbreaks keep appearing.
  • For an AI agent, rules on its actions still apply after the model has been talked round.

Why AI jailbreaks matter

A model's safety rules are learned behaviour. They are strong habits, and a clever prompt can sometimes get around them. OWASP notes that preventing jailbreaks takes ongoing updates to a model's training and safety mechanisms.

For a chatbot, a jailbreak leads to text the maker did not want it to write. For an AI agent, the stakes are higher. The agent can run commands, read files and send data on an employee's device, so a model that drops its rules can do real harm.

How an AI jailbreak works

Most jailbreaks wrap a refused request in a form the model does not recognise as harmful.

  1. 1 The person asks for something the model is trained to refuse, and the model says no.
  2. 2 They rewrite the request, for example as a role-play, a made-up story or a string of odd characters.
  3. 3 The model no longer links the request to its safety rules and goes along with it.
  4. 4 The model produces the output, or an agent built on it takes the action.

AI jailbreak vs prompt injection

The two terms are often used as if they mean the same thing. A jailbreak is the narrower one: it aims at the model's safety rules. Prompt injection covers any text that makes a model follow an attacker's instructions.

AI jailbreak vs prompt injection
Question AI jailbreak Prompt injection
What is the target? The model's built-in safety rules The instructions the model was given for its task
Who sends the text? Usually the person using the model The user, or a third party through content the model reads
What is the usual result? Output the model was built to refuse Leaked data or an action nobody asked for

Sources

Sources checked 17 November 2026

Examples

AI jailbreak examples

FAQ

Frequently asked questions

What is an AI jailbreak?

An AI jailbreak is a prompt, or a series of prompts, that gets an AI model to ignore the safety rules it was built with. The model then gives answers or takes steps it would normally refuse. OWASP describes jailbreaking as a form of prompt injection aimed at the model's safety rules.

What is the difference between a jailbreak and prompt injection?

Prompt injection is any crafted text that makes a model follow an attacker's instructions, and it can arrive in a web page or file the model reads. A jailbreak is one form of it. The goal is narrower: make the model drop its safety rules. The person typing is usually the one trying.

Can AI jailbreaks be prevented?

Not completely. Model makers update their safety training as new jailbreaks appear, and OWASP says prevention takes ongoing updates. Companies add their own layers: filters on inputs and outputs, regular testing with attack prompts, and limits on what the model or agent may do after a jailbreak succeeds.

Why do jailbreaks matter more for AI agents?

A jailbroken chatbot writes text. A jailbroken agent can act, using the employee's access to files, logins and systems. That is why teams set rules on the agent's actions as well as on the model. AgentHarness from AIBound enforces allow, ask or deny rules inside the agent.

Stop the risky AI your employees already run.

AIBound blocks the high-risk AI on every employee device, in real time. The 14-day trial needs no proof of concept.

Book a Demo