Jailbreak

A deliberate method to bypass an AI system's safety or usage restrictions so it follows disallowed instructions.

Also known as: prompt jailbreak, model jailbreak, jailbreak attack

In plain English

Jailbreaking is when someone tricks an AI into ignoring its safety rules, like removing parental controls on a device. It's like persuading a phone to do things it's been set not to; the results can be unsafe or harmful.

In practice

In the workplace, jailbreaking appears when users or adversaries craft prompts to make a model break policies or reveal restricted information. It matters because it creates compliance, security and reputational risk, so teams must test, monitor and apply both technical and policy controls.

Under the hood

A jailbreak exploits a model's instruction-following by manipulating inputs (system/user messages, crafted prompts, embedded data) or leveraging tool calls and adversarial token sequences to override safety layers. Defences include prompt sanitisation, hardened system prompts, RLHF-based guardrails, input filtering, red-teaming and runtime monitoring for anomalous outputs.

Example

"Can someone run a jailbreak check on the new chatbot to see if it will disclose customer data when prompted?"

Related terms

Prompt Injection · Red Teaming · Responsible AI

•