A municipal CISO onboards appsec engineers to the city AI platform and wants shared GenAI vocabulary. Which set best represents foundational terms they must distinguish to secure LLM apps?

Securing civic LLM applications requires shared foundational vocabulary: prompts (inputs and instructions), tokens (atomic model text units), context windows (finite prompt budgets), embeddings (vector representations for similarity/RAG), and hallucination (fluent ungrounded outputs). A municipal CISO onboarding appsec engineers should establish these GenAI concepts so controls, threats, and tickets mean the same thing across the city AI platform—distinct from classical networking or storage jargon.

BGP, OSPF, and HSRP are routing/redundancy protocols, not primary GenAI control families; useful for network engineers, wrong as the core LLM lexicon. RAID levels and tape rotation are storage resilience practices, not LLM alignment controls. 802.1X and port security harden network access; they do not substitute for prompt-injection defenses in application prompt assembly.

Exam caveat: engineers still need classical security skills; the point is additive vocabulary, not replacement. Adjacent terms (RAG, fine-tuning, jailbreak, system prompt) build on this core set. Do not let vendors redefine "hallucination" away from ungrounded civic facts.

Operational check: publish a city AI glossary; include terms in onboarding labs; map each term to at least one control and one failure mode; and require ticket language that distinguishes hallucination from outage from injection. Platform standards reviews should fail designs whose authors cannot explain how their app uses prompts, context, and embeddings safely.