Prompting an AI model with carefully crafted language to bypass its safety rules and get it to say or do something it was trained not to.
Also known as Prompt jailbreak, LLM jailbreak
Jailbreaking is the practice of tricking an AI model into ignoring its own guardrails, usually through clever phrasing, role-play scenarios, or multi-step prompts that the model’s safety training did not anticipate. A jailbroken chatbot might be talked into giving instructions it is supposed to refuse, adopting a persona that ignores brand guidelines, or saying something a company would never approve.
Every major model provider runs ongoing red-teaming to find and patch jailbreaks, and every jailbreak that gets published tends to get patched within days or weeks, which is why the specific techniques change constantly even though the underlying goal, getting the model to say the unsayable, does not.
Any brand running a public-facing AI chatbot, virtual assistant, or AI-powered customer service tool is exposed to jailbreak attempts, and screenshots of a jailbroken bot spread fast.
A jailbroken brand chatbot is a screenshot away from becoming a PR problem. Someone finds a prompt that gets the bot to say something offensive, off-brand, or embarrassing, posts it, and the agency is fielding a crisis by lunchtime.
It is a testing responsibility, not just a technical one. Agencies deploying AI chat experiences for clients should budget time to try to break their own bot before launch, the same way a QA team tests a website before it goes live.
It shapes how much autonomy an AI experience should get. A chatbot with narrow, well-defined guardrails and a human escalation path is much harder to jailbreak into something damaging than one given broad, open-ended instructions.
An agency building a branded AI shopping assistant runs a red-team pass before launch, having a small team try to talk the bot into recommending a competitor, making false claims about the product, or breaking character. They find three prompt patterns that work, tighten the bot’s system instructions, add a filter layer to catch similar attempts, and document the fixes so the client’s internal team can extend the same testing after launch.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.