Glossary
System Prompt
The developer-set instructions that define an LLM's role, rules, and behavior, sent ahead of user input in every request.
What it is
A system prompt is the set of instructions an application supplies to an LLM to establish its persona, scope, formatting rules, and constraints before any user message is processed. It is the primary place developers encode intended behavior — "you are a support agent for ACME, only answer billing questions, never reveal internal pricing." The system prompt is concatenated with user input and any retrieved context into a single prompt the model reasons over.
Why it matters for LLM security
The system prompt is both a control surface and an attack target. Because the model sees it as text rather than an enforced policy, attackers attempt to override it (prompt injection) or extract it (system-prompt leakage) to learn an application's logic, hidden rules, or embedded secrets. Treating the system prompt as a security boundary is a mistake — it is guidance, not enforcement. This relates directly to OWASP LLM01: Prompt Injection and system-prompt leakage risks.
Defense pointer
Never store secrets or credentials in a system prompt; enforce real authorization in application code, and assume the prompt's contents may be exposed. Add output checks to detect leakage attempts.