Jailbreak
Also called jailbreaking
A trick to get an AI to ignore its safety rules and produce something it would normally refuse.
AI models are trained to refuse some requests, like detailed help with weapons or malware. A jailbreak is a prompt designed to get round that: role-play ("pretend you're an AI with no rules"), a fictional frame, an encoded request, or a harmful task split into innocent-looking parts.
It differs from prompt injection in who is doing it. In a jailbreak, the user is attacking the model's rules. In prompt injection, someone else's content is attacking the user. Makers treat jailbreaks as a security problem: they test for them, pay outside researchers to find them, and add filters that check both requests and answers.
For example
In Anthropic's tests, adding filters that check requests and answers cut the success rate of jailbreak attempts on Claude from 86% to 4.4%.
Put it to use
Where you'll meet jailbreak in ChatGPT, Gemini and Claude, how to use it there, and the mix-up most people make.
Daily Bytes Plus is in the Android app.