A veterans services office notes that attackers can hide instructions inside user text to override a benefits chatbot’s rules. How should they distinguish that threat?
Select an answer to reveal the explanation.
Short Explanation
SQL injection sneaks commands into a database query; prompt injection sneaks commands into the model’s instructions. Same family of ‘don’t trust user input,’ different door on the building.
Full Explanation
When attackers hide instructions inside user text to override a veterans services benefits chatbot’s rules, staff should recognize that threat as prompt injection—an AI-specific threat distinct from classic SQL injection. Prompt injection targets the model’s instruction-following and tool-use behavior: adversarial content in a claim note, pasted email, or chat turn tries to make the assistant ignore system policy, reveal restricted guidance, or call tools inappropriately. Defenses center on input handling, privilege separation for tools, output filtering, and not treating retrieved or user text as trusted system instructions.
Treating it as identical to SQL injection and only hardening the relational database misses the attack surface: even a perfectly parameterized database can sit behind a model that obeys hostile natural-language instructions. Ignoring the threat because chatbots cannot be influenced by user-supplied text is factually wrong; instruction-following models are specifically sensitive to crafted user content. Classifying it as a physical badge-cloning problem for the lobby kiosk confuses endpoint physical security with a software/AI control-plane abuse pattern.
Exam caveat: prompt injection and SQL injection share a “don’t trust user input” family resemblance, but exam answers require naming the AI-specific mechanism when the stem describes overriding model rules via text. Operational check: separate system prompts from untrusted content, constrain tools, red-team the benefits bot with injection payloads, and monitor for policy-bypass attempts.