Simple definition
An AI jailbreak attempts to bypass an assistant's instructions or safeguards through manipulated inputs.
Purpose and use
An AI jailbreak attempts to bypass intended behavioral or content restrictions of a model or application.
How does it work?
It seeks output that should be disallowed, potentially using wording or context that misleads control mechanisms.
Example and verification
In authorized evaluation, define the prohibited outcome and separately inspect refusals, produced answers and possible tool actions.
Points to watch
One failed attempt does not prove general resistance. An apparently compliant answer can also follow an unwanted action already executed by a tool-enabled system.
In short
Evaluate observable effects across representative scenarios and state the limits of the conclusion.