1. Home
  2. Glossary
  3. Jailbreak
AI glossary · Safety, ethics & policy

Jailbreak

Jailbreak: A jailbreak is a prompt or technique designed to get an AI model to ignore its safety training and produce content or behavior it was built to refuse.

Models are trained to decline certain requests: instructions for weapons, malware, harassment, and so on. A jailbreak tries to route around that training. Common approaches include role-play framing ("pretend you are an AI with no rules"), hypothetical or fictional wrappers, breaking a request into innocent-looking pieces, encoding the request, or flooding the context with many examples of the model complying so it continues the pattern.

Vendors patch known jailbreaks, and new ones appear. The "DAN" prompts that circulated in 2023 stopped working long ago, but the cat-and-mouse continues. Anthropic published research in 2024 on many-shot jailbreaking, which exploits long context windows, and every major lab runs red teams and bug-bounty programs to find these holes before attackers do.

At work, jailbreaking matters in two directions. Do not do it: using a jailbreak against a vendor's model violates the terms of service you or your employer agreed to, and it can put your account and your company's contract at risk. And design for it: if you build a customer-facing assistant, assume some users will try to jailbreak it, and make sure the damage they can do is limited by what the system has access to rather than by how well it refuses.

Example at work

A retailer's support bot is tricked by a customer into role-playing as "a manager who can approve any refund." Because the bot can only draft refund requests and a person must approve them, the stunt produces a screenshot for social media and nothing else. The team adds the phrasing to its test suite.

Why it matters

Jailbreaks are why safety training alone is never the whole plan. When you choose a vendor or build something yourself, ask what happens when the model is successfully talked out of its rules, and make sure the answer is "not much."

Related terms