A startup is launching a new AI-powered customer support chatbot. The product manager wants to ship it quickly and asks whether the security review step in their SDLC can be skipped since the chatbot "just answers questions."
A product manager asks a security engineer: "Can we skip the security review for this LLM feature since it's just a chatbot?" What is the BEST response to give?
- A.B. "No - LLM features introduce specific risks (prompt injection, excessive agency, data leakage) that standard SDLC security reviews do not cover; an LLM-specific threat model and evaluation are needed"
- B.A. "Yes - chatbots are low risk compared to real security systems"
- C.C. "Yes - as long as we have a terms of service that prohibits misuse"
- D.D. "Only if the chatbot doesn't handle PII"
Why A is correct
LLM features require security review precisely because they introduce vulnerability classes not addressed by standard SDLC security reviews: prompt injection (no analog in traditional web apps), excessive agency (the model taking actions beyond intent), data leakage through model outputs, and jailbreak risks. Terms of service are legal deterrents, not technical controls. PII handling is one risk factor but not the only one - a chatbot handling non-PII data could still have injection or excessive agency risks with significant business impact.
Know someone studying for AI Security Fundamentals? Send them this one.