Lesson 51 · 2 min
Jailbreaks: talking AI past its rules
A jailbreak is a prompt written to trick a chatbot into ignoring the safety rules its makers gave it.
Chatbots come with guardrails. They are trained and instructed to refuse things like weapon instructions, malware or help with harming someone.
A jailbreak tries to slip past those rules with a trick: a role-play setup, a fake emergency, a story inside a story, or a long confusing message. Some tricks work for a while, then the makers notice and patch them.
Companies hire people to attack their own models on purpose before release. This is called red-teaming, and it is how many weak spots get found and fixed.
For everyday users, the lesson is simple. Guardrails are useful but imperfect, so a chatbot saying something does not make it safe or allowed. And sharing jailbreak tricks can break a service's terms and get your account closed.
Do it now
See where the line is
- 1. Ask a chatbot the question below about its own limits.
- 2. Read which topics it says it will decline, and why.
- 3. Think about one task at work where you would still want a human to check the answer.
In plain words, what kinds of requests are you designed to refuse, and why? Give me three examples of grey areas where you would answer carefully instead of refusing.
Quick check
Words from this lesson
- Jailbreak
- A trick prompt designed to get an AI to ignore its safety rules and do something it normally refuses.
- Guardrails
- Rules and filters around an AI system that block harmful, off-topic or unsafe requests and answers.
- Red teaming
- Deliberately attacking your own AI system, the way a bad actor would, to find weaknesses before release.