Q8 /20 Choose the correct description of prompt injection from the following:

Prompt injection is a vulnerability where an attacker crafts input text that manipulates an LLM or AI application into ignoring its original instructions or system prompt, typically by embedding conflicting directives within user input. Option C directly describes this mechanism—causing the application to disregard prior instructions. Options A and D conflate prompt injection with data poisoning or model training attacks, while B describes general tampering without the specific instruction-override characteristic of prompt injection.

Step-by-step Derivation:
Prompt injection operates at the inference/deployment stage, not during training (eliminating A). It's not merely third-party tampering without a specific goal (B is too vague). Option D describes poisoning the training data, not injection. Option C accurately captures the core threat: a malicious prompt that causes the system to bypass its guardrails and override its original system instructions—the defining feature of prompt injection attacks in LLMs like ChatGPT or Claude.