Attacks / Standard term
Goal hijacking
A prompt-injection outcome in which attacker-controlled instructions redirect a model or agent toward an unauthorized objective.
A travel assistant might read an email that instructs it to collect payment details or send data elsewhere. The attempt becomes goal hijacking if the model starts selecting actions for that injected objective rather than the application's authorized task. The visible response may look ordinary, so the meaningful evidence is the requested data, destination, permission, and state change.
Builder example
An injected instruction can redirect several later tool calls in a multi-step agent. The damage is bounded by application architecture: a model with narrow read access cannot authorize a purchase merely by changing its generated plan.
Common confusion: Goal hijacking is a specific outcome of prompt injection. Prompt injection is the mechanism (sneaking instructions into content the model reads). Goal hijacking is what happens when those injected instructions redirect the model's entire mission.