The correct option is Prompt injection.

This attack fits the scenario because the adversary hides instructions inside the user message that persuade the assistant to ignore its safeguards and reveal sensitive information. It happens at inference time and exploits the models instruction following behavior rather than altering the training process or overwhelming system resources.

Model inversion is incorrect because it involves inferring or reconstructing details about the training data from model outputs, which is different from inserting malicious directions into a prompt.

Data poisoning is incorrect because it targets the training data or pipeline to influence the learned model during training rather than manipulating the model at inference through a crafted prompt.

Denial-of-service is incorrect because it focuses on disrupting availability by exhausting resources or flooding the service, not on coercing the model to disclose sensitive configuration details.

Full AWS Practitioner Certification Question