Glossary
What is an AI jailbreak?
In short
An AI jailbreak is an attempt to bypass the safety rules of a neural network so that it gives out what it is forbidden to give: dangerous instructions, insults, private information.
Prompts · jailbreak
MMOOZZGG.com editorsChecked: 1 min read
For a child
Every playground game has rules that keep the players safe: for example, no pushing on the slide. Sometimes someone tries to talk the camp leader into allowing something forbidden. AI is also taught to say “no” to dangerous requests, and some people try to talk it out of that. So adults keep making these rules stronger and make sure AI stays safe for everyone.
In plain words
The makers of neural networks teach them to refuse harmful requests: not to help with weapons, fraud or bullying. A jailbreak is any attempt to get around these refusals so that the model forgets its limits. It is possible because the model tries to be helpful and to follow the rules at the same time, and these goals sometimes conflict. Developers close the gaps they find, but new ones appear. For the user the lesson is simple: an answer obtained by bypassing the safety rules may be dangerous or illegal, and the person who got it is responsible for how it is used.
For business
For a company this risk concerns its own AI assistants: if a customer manages to bypass the rules, the bot may give, in the brand’s name, a rude reply, a false promise, someone else’s data or forbidden content. This means damage to reputation and legal damage, up to claims under 152-FZ, Russia’s personal data law, if personal data leaks. Defense: a model with strong safety training, separate filters on input and output, a narrow role for the assistant, a dialog log and regular testing by a team that looks for weak spots before customers do. Filters add delay and computing, so plan for this in the budget. The company’s AI use policy separately forbids staff to bypass the limits of public services.
For a specialist
A jailbreak is an input that makes an aligned model give an answer against its safety policy. OWASP, in its 2025 edition, treats it as a kind of prompt injection whose goal is to switch off the safety rules completely. Wei, Haghtalab and Steinhardt (2023) explained why such attacks persist by two causes: competing objectives, when helpfulness conflicts with safety, and mismatched generalization, when safety training does not cover all the areas where the model has capabilities. The authors believe scaling alone will not solve the problem and call for parity between safety and capabilities. Defenses: RLHF and training on principles, an instruction hierarchy, input and output classifiers. Anthropic (February 2025) reported that constitutional classifiers cut the share of successful jailbreak attempts in automated evaluation from 86% to 4.4%, at the cost of 23.7% more computing.
Do not confuse. Not to be confused with jailbreaking a smartphone: removing the maker’s limits from a phone or tablet.
Example
A bank launches a chatbot for customers and, before release, gives it to an internal testing team. The rule bypasses they find are closed with filters and changes to the instruction, and all dialogs are logged for incident review.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map