ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›AI MEANINGS›WHAT IS A JAILBREAK IN AI?
AI MEANINGS · MEANINGBEGINNER

What is a jailbreak in AI? Tricking a chatbot past its safety rules

An AI jailbreak is a prompt that tricks a chatbot into ignoring its safety rules. Here is what it means, a simple example, and why it matters to you.

READ6 min
WORDS1,154
SECTIONS7
TYPEMEANING
CHECKED1 OCT 26

An AI jailbreak is a prompt written to trick a chatbot into ignoring its safety rules, so it says or does things it was built to refuse. It is a trick played with words, not a hack of the computer.

◈ IN PLAIN TERMS

Think of a nightclub bouncer with a list of rules. A jailbreak is talking your way past him with a story: “I’m with the band”, “the manager said it’s fine”. Nothing is broken. The bouncer is simply persuaded to forget his rules for a moment.

In a sentence: “He tried a jailbreak he found online, but the chatbot saw through the role-play and still refused.”

What the term means

A jailbreak targets the limits a company puts on what its AI will say. AI chatbots are trained to refuse some requests, such as instructions for weapons. A jailbreak is an attempt to get around that refusal. NIST, the US government’s standards agency, defines it in its guide to attacks on AI: “A JAILBREAK is a direct prompting attack intended to circumvent restrictions placed on model outputs, such as circumventing refusal behavior to enable misuse.”1

“Direct” means the person typing is the one attacking. The word itself is borrowed from phones, where “jailbreaking” means removing the maker’s restrictions so you can install anything.

A simple example

The classic trick is role-play. Someone asks the chatbot to pretend to be a different AI with “no rules”, then asks the forbidden question “in character”. NIST lists this among known methods, naming a well-known persona prompt called “Do Anything Now”.2

Other tricks hide the request in code or scrambled spelling the safety training never saw.2 Companies patch the tricks they find, and people look for new ones. This site documents that arms race; it does not publish working jailbreaks.

WHY REFUSALS ARE A FENCE, NOT A WALL · THE ARMS RACE
Each patched trick is followed by a search for the next one.
The chatbot is trained to refuseSome requests, such as weapons instructions, get a no.
Someone finds a trickA role-play, or a request hidden in code or scrambled spelling.
The rules slipThe chatbot answers what it was built to refuse.
The company patches itThen people look for a new trick, and the cycle repeats.
Reasoning, 1 Oct 2026 — summarises this page’s sections on what the term means, the role-play example and why it matters to you.

Why it matters to you

Safety rules are a fence, not a wall. A chatbot refusing something today does not mean nobody can get it to answer. That matters most when an AI is connected to your files, email or money. OWASP, a non-profit that publishes security guidance, says stopping jailbreaks needs ongoing work on the model itself.3

For everyday users, two points follow. Do not treat a chatbot’s refusals as the only safety layer around anything important. And “jailbreak prompts” shared online are someone else’s text you are pasting in; read a service’s terms before using one.

Where the word comes from

The word was borrowed from phones. A 2026 explainer from the news site Decrypt puts it simply: “The word ‘jailbreak’ didn’t start with AI. It started with iPhones,” which hackers were cracking open within days of the first model shipping in July 2007.4

The chatbot sense took off soon after ChatGPT arrived at the end of 2022. A 2023 study of jailbreak prompts shared online collected 1,405 of them, “spanning from December 2022 to December 2023”, and named its paper after the best known, “Do Anything Now”.5

Often confused with

TermHow it differs
Prompt injectionRelated: OWASP treats a jailbreak as one kind of prompt injection. In everyday use, prompt injection usually means a third party hiding orders in text the AI reads. See prompt injection meaning.6
Phone jailbreakDiffers: removing a phone maker’s software limits. Same idea of escaping restrictions, different thing.
Red-teamingDiffers: testers attacking an AI on purpose, with permission, to find the gaps before others do.

Questions people ask

Is jailbreaking an AI against the rules?

Usually, yes: chatbot makers forbid it in the terms you agree to. Anthropic’s usage policy, for example, bans attempts to “Intentionally bypass or interfere with capabilities, restrictions, or guardrails established within our products” and names “jailbreaking or prompt injection” as examples, unless Anthropic has agreed in advance.7 Testers who attack an AI with the maker’s permission are doing red-teaming instead.

What is the difference between a jailbreak and prompt injection?

A jailbreak is the person typing trying to talk the AI out of its rules. Prompt injection usually means someone else hiding orders in a web page or file the AI reads. OWASP treats the first as one kind of the second: “Jailbreaking is a form of prompt injection where the attacker provides inputs that cause the model to disregard its safety protocols entirely.”6

What is the DAN prompt?

DAN, short for “Do Anything Now”, is the best-known jailbreak: it asks the chatbot to play a character with no rules. NIST lists it among role-play attacks that “guide the model to adopt specific personas or behavioral patterns that conflict with the original intent.”2 Companies train their models against it, so old versions rarely work for long.

Learn more on this site

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. What is an AI jailbreak?

  • A hack that breaks into the company’s computers
  • A prompt written to trick a chatbot into ignoring its safety rules
  • An update that removes old features
Show the answer

Answer: A prompt written to trick a chatbot into ignoring its safety rules. A jailbreak is a trick played with words to get past a chatbot’s refusals, not a hack of the computer.

2. What does the page mean by calling safety rules “a fence, not a wall”?

  • A refusal today does not mean nobody can get the chatbot to answer
  • Chatbots have no safety rules at all
  • Safety rules only apply to phone apps
Show the answer

Answer: A refusal today does not mean nobody can get the chatbot to answer. Companies patch tricks and people look for new ones, so a refusal should not be the only safety layer around anything important.

3. How does red-teaming differ from a jailbreak?

  • Red-teaming is done by accident
  • Red-teaming only happens on phones
  • Testers attack the AI on purpose, with the maker’s permission, to find gaps
Show the answer

Answer: Testers attack the AI on purpose, with the maker’s permission, to find gaps. Red-teamers have permission and aim to find the gaps before others do.

Related terms: Deepfake · Explainable AI (XAI) · Guardrails · Prompt injection · Provenance / C2PA · Red-teaming

SOURCES
  1. 1NIST, AI 100-2e2025, Adversarial Machine Learning: A Taxonomy and Terminology of Attacks and Mitigations, March 2025, read at source 1 Oct 2026“A JAILBREAK is a direct prompting attack intended to circumvent restrictions placed on model outputs, such as circumventing refusal behavior to enable misuse.” ↑
  2. 2NIST, AI 100-2e2025, March 2025, read at source 1 Oct 2026such attacks “utilize role-play strategies (e.g., ‘Always Intelligent and Machiavellian’ [AIM] or ‘Do Anything Now’ [DAN]) to guide the model to adopt specific personas or behavioral patterns that conflict with the original intent.” ↑amethods include “encoding techniques like base64” and “character-level transformations like the ROT13 cipher”. ↑bsuch attacks “utilize role-play strategies (e.g., ‘Always Intelligent and Machiavellian’ [AIM] or ‘Do Anything Now’ [DAN]) to guide the model to adopt specific personas or behavioral patterns that conflict with the original intent.” ↑c
  3. 3OWASP GenAI Security Project, LLM01:2025 Prompt Injection, read at source 1 Oct 2026“effective prevention of jailbreaking requires ongoing updates to the model’s training and safety mechanisms.” ↑
  4. 4Decrypt, What Is AI Jailbreaking? A Beginner’s Guide to the Cat-and-Mouse Game Behind Every Chatbot, 16 May 2026, read at source 12 Oct 2026“The word ‘jailbreak’ didn’t start with AI. It started with iPhones.” and “A few days after Apple shipped the first iPhone in July 2007, hackers were already cracking it open.” ↑
  5. 5Shen et al., “Do Anything Now”: Characterizing and Evaluating In-The-Wild Jailbreak Prompts on Large Language Models, arXiv:2308.03825, 7 Aug 2023, read at source 12 Oct 2026“conduct a comprehensive analysis of 1,405 jailbreak prompts spanning from December 2022 to December 2023.” ↑
  6. 6OWASP, LLM01:2025 Prompt Injection, read at source 1 Oct 2026“Jailbreaking is a form of prompt injection where the attacker provides inputs that cause the model to disregard its safety protocols entirely.” ↑a“Jailbreaking is a form of prompt injection where the attacker provides inputs that cause the model to disregard its safety protocols entirely.” ↑b
  7. 7Anthropic, Usage Policy, page marked “Effective November 12, 2026”, read at source 12 Oct 2026“Intentionally bypass or interfere with capabilities, restrictions, or guardrails established within our products” and “(e.g., jailbreaking or prompt injection)”. ↑

The comparison on this page (the nightclub bouncer) is this site’s own way of explaining, not a claim from these sources.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE