A transit authority compares attacks on a rider-alert LLM with classic SQL injection on the same website. What distinction should the offensive AI lead emphasize?
Select an answer to reveal the explanation.
Short Explanation
SQL injection is sneaking malicious orders into the kitchen ticket printer. AI input abuse is talking the front-desk agent into ignoring the house rules or calling the wrong tool—same building, different door.
Full Explanation
On a transit rider-alert site, classic SQL injection and AI input attacks may coexist, but they abuse different interfaces. SQL injection manipulates structured query languages against the database. AI input attacks manipulate natural-language instructions, retrieved context, or agent tool selection so the model ignores policy or calls the wrong capability. Offensive AI leads must classify findings correctly so remediation lands on parameterized queries versus prompt/context controls, output handling, and tool authorization.
Claiming parameterized queries and strict input validation reduce prompt manipulation in the rider-alert model is wrong: those controls harden database query interfaces, but they do not stop natural-language instructions, context, or tool selection from steering the LLM. Claiming web application firewall signatures for SQL injection provide coverage for malicious instructions sent to the rider-alert model is wrong: WAF SQL patterns target database query syntax, not the natural-language prompts or context that abuse an LLM. Claiming API-level encryption reduces AI input attacks confuses transport security with application-level instruction integrity—encrypting the channel does not filter or validate what instructions or context reach the model, so instruction-level manipulation is not meaningfully reduced by encryption alone.
Exam caveat: distinguish query-language injection from AI instruction/context/tool abuse; TLS is not an AI input control. Operational check: on the rider-alert stack, separate findings into DB query flaws versus prompt/RAG/tool steering; verify parameterized queries and distinct AI controls (prompt filters, retrieval integrity, tool auth); retest each plane under authorized RoE.