Prompt injection

Technique inserting adversarial instructions into a model's input (often via user data or tool output) to override the system prompt's safety constraints.