Select the correct prompt injection attack from the following list for each definition.

Prompt injection attacks attempt to manipulate large language models into disregarding intended instructions or revealing protected information.

Fake completion involves supplying partially completed responses that influence subsequent model output. Extracting conversation history attempts to retrieve previous interactions that might contain confidential information. Ignoring the prompt template directly instructs the model to disregard its established behavioral constraints.

Extracting the prompt template exposes hidden instructions or formatting structures. AWS explains that attackers can subsequently imitate exposed template tags to introduce malicious instructions.

The fourth description specifically concerns template-tag spoofing, which AWS identifies as a potential consequence of prompt-template extraction rather than extraction itself. However, option B is the intended association among the available choices.

AWS recommends layered protection against prompt injection, including input validation, access controls, and appropriate guardrails.