---
title: Jailbreak
description: 'An input strategy intended to make a model bypass or disregard its trained or instructed safety restrictions.'
canonical_url: 'https://darkfactory.dev/glossary/jailbreak'
markdown_url: 'https://darkfactory.dev/glossary/jailbreak.md'
collection: glossary
date_published: '2026-08-03T00:00:00-04:00'
date_modified: '2026-08-26T00:00:00-04:00'
---

# Jailbreak


## Definition

A jailbreak is an input strategy intended to make a model bypass its trained or instructed safety restrictions. It may use role-play, encoded requests, fictional framing, long conversations, conflicting instructions, or adversarial suffixes to elicit behavior the model would normally refuse.

Jailbreak testing tells you about a model's resistance to prohibited requests, but a refusal is not an authorization system. Model behavior can change across versions and inputs, and an attacker only needs one useful failure. Applications that can read secrets or take actions still need access control, constrained tools, monitoring, and limits enforced outside the model.

## Distinguish it from nearby terms

Jailbreaking targets model safety restrictions. Prompt injection more broadly redirects an application's intended behavior, including attacks on tools, data, and workflow instructions. A direct prompt injection can also be a jailbreak when its goal is to evade a refusal.

## Check your understanding

A chatbot refuses instructions for an unsafe chemical process but has no tools or private data. That is primarily a model-safety problem. If the same input makes a laboratory agent operate equipment, the application also has an authorization and containment failure.

## Also called

jailbreaking

## Related factory areas

- [Security, privacy & software supply chain](https://darkfactory.dev/factory/security)

## Evidence and further reading

- [NIST AI 100-2: Adversarial Machine Learning](https://csrc.nist.gov/pubs/ai/100/2/e2025/final)
- [OWASP GenAI Security Glossary](https://genai.owasp.org/glossary/)