Prompt Injection is the correct risk because the attacker places crafted instructions in input that cause the language model to disregard or alter its intended behavior. The instructions may be directly visible to a user or hidden inside content that the application supplies to the model. OWASP defines LLM01 Prompt Injection as user prompts changing model behavior or output in unintended ways and notes that the influencing content need not be human-visible. Output truncation concerns incomplete responses, not instruction manipulation. System prompt leakage is the exposure of hidden system instructions; it can be a consequence of prompt injection but is not the injection technique itself. Data exfiltration describes unauthorized information removal and likewise can be an outcome, rather than the vulnerability identified by the stem.

================