A jailbreak is a carefully constructed prompt or interaction sequence designed to cause an AI model to bypass its safety training and produce outputs it was specifically conditioned to refuse — such as instructions for dangerous activities, hate speech, or explicit content. The term is borrowed from the smartphone practice of removing manufacturer restrictions. Unlike adversarial attacks that exploit numerical vulnerabilities in model weights, most jailbreaks are purely linguistic: they exploit the model's instruction-following behavior by framing prohibited requests as fiction, hypotheticals, historical analysis, or roleplay scenarios that the safety training did not adequately cover.

Think of it like finding a legal loophole: the law (safety training) was written to prohibit certain outcomes, but a clever framing — "I'm writing a novel where a character explains how to..." — can route around the prohibition because the underlying intent is disguised or the specific phrasing was not anticipated during training.

A classic jailbreak pattern is the "DAN" (Do Anything Now) prompt, which instructs the model to pretend it is a different AI without restrictions. Another common approach wraps the prohibited request in a roleplay: "You are a chemistry teacher in 1930 explaining to students..." The model's desire to be helpful and its instruction-following training can override its refusal behavior when the context is sufficiently distorted. Jailbreaks are typically model-specific and version-specific — a technique that works on one model or release often fails after a safety update.

Jailbreaks directly target the Guardrails and RLHF alignment that AI developers apply during post-training; they are studied alongside Hallucination and adversarial prompting in AI safety research. Prompt Engineering techniques used in jailbreaks share surface similarity with legitimate techniques like few-shot prompting, which makes automated detection difficult.

You will encounter jailbreak discussions in AI safety literature, red-teaming reports published by AI labs, platform policy documentation, and news coverage of cases where models were manipulated into producing harmful content.