Skip to main content

COMPEL Glossary / system-prompt-leakage

System prompt leakage

Extraction of an LLM feature's hidden system prompt and structural instructions through crafted user input.

What this means in practice

Creates intellectual-property exposure (leaked instructions reveal product design) and security exposure (leaked scaffolding reveals tool contracts and guardrail boundaries).

Synonyms

system prompt extraction , LLM07

See also

  • Guardrail — A control placed between the user or environment and an LLM that blocks, rewrites, or classifies content at one of four architectural layers: input filter, policy filter, output filter, or tool-call validator.
  • Model and prompt registry — A versioned inventory of models, system prompts, retrieval sources, and guardrails deployed in production.