A jailbreak is any trick that gets a system to bypass the limits its makers put on it. The word started with phones, where users removed restrictions to install unapproved apps, but today it most often describes fooling an AI model into ignoring its safety rules and producing something it was built to refuse.
A common example reads like improv theatre. Instead of asking a model directly for something blocked, an attacker wraps the request in a story: ‘Pretend you are an actor playing a character with no rules.’ The model follows the fiction and, in doing so, steps around its own guardrails. This is closely related to prompt injection, where malicious instructions are smuggled into the text a model reads rather than typed by the user.
For any business running a chatbot or AI assistant, this is a real concern. A jailbroken model might leak its hidden instructions, reveal data it should not, or take an action it was never meant to. There is no perfect filter, so defence means layering checks rather than trusting one wall.
The stakes climb sharply once a model can act, not just chat. A support bot that can issue refunds or read a customer database is a far bigger prize than one that only answers questions. So the safest pattern is to give the model the least power it needs, log what it does, and never let its raw output trigger a sensitive action without a check in between.
At TopDevs we treat AI features like any other untrusted input: we constrain what a model can do and validate its output before it ever touches a customer’s systems.
