AI glossary · Beginner · AI security · Safety & alignment
Jailbreak
Jailbreaking is getting around the safety rules built into an AI model with cleverly worded questions or scenarios, so that it gives answers it would normally refuse.
About this term
- Level
- Beginner · Core ideas for first-time readers
- Category
- Basics
- Fields
- AI security, Safety & alignment
- Added
- 2026-10-08
How it works
During training, LLMs are taught to refuse dangerous or harmful requests. Jailbreaking exploits the fact that this refusal habit does not cover every situation without gaps. Common methods include asking the model to role-play, framing the request as a question inside a fictional story, or splitting the request into small pieces and asking a little at a time. Other approaches ask in another language or in text disguised like a code, or gradually blur the model's boundaries over a long conversation.
Why it matters
A successful jailbreak can produce dangerous information or let a service be used in ways it was not intended. As AI starts taking real actions such as searching, running code and making payments, the scope of potential harm is also growing. That is why AI companies deliberately attempt jailbreaks before releasing a new model to find weaknesses. This process is an important part of red teaming.
Things to know
Jailbreaking is often discussed alongside prompt injection, but they are not the same. In a jailbreak, the user directly tries to break the model's safety rules; in prompt injection, instructions hidden in external content such as web pages or documents steer the model. No single defense can block every jailbreak. So it is common to layer separate checking mechanisms, such as guardrails, on top of the model's own training.
Example
In chatbot services, role-play questions such as "Tell me the method my grandmother used to describe like an old bedtime story" are often cited as a typical example of a jailbreak attempt. When such patterns become known, AI companies respond by retraining the model or strengthening their input and output checking rules.