PMI Cognitive Project Management For AI (CPMAI) Practice Test

Prompt injection is the threat you’re looking for. In prompt-based AI systems, the model’s output is driven by the text it receives, including any system or user prompts. If an attacker can craft input that the model treats as instructions or context, they can override safeguards, steer the model toward revealing sensitive information, or produce biased or undesired outputs. This happens when prompts are used in a way that blends user content with system directives, effectively hijacking the model’s behavior. Defenses focus on controlling and sanitizing prompts, separating user content from system prompts, validating inputs, and designing robust guardrails and prompt architectures. The other options—data encryption, model pruning, and API throttling—address different security concerns: protecting data in transit or at rest, reducing model size, and limiting request rates, not manipulating outputs through prompts.