An AI jailbreak is a prompt or technique designed to get an AI system to ignore restrictions. It may ask the model to roleplay, reveal hidden instructions, or provide disallowed content.
In practice
Jailbreaks target the behavior layer of an AI system. They are common in public chatbots and can also affect internal tools if users or documents include adversarial instructions.
What to watch
Jailbreak resistance requires layered defenses. Policy prompts help, but apps also need monitoring, filtering, permissions, and careful tool design.