.svg)
What is OWASP LLM07 (2025) System Prompt Leakage?
System prompt leakage is when crafted inputs coax the model into revealing instructions that were meant to stay private - safety rules, business logic, tool definitions, or internal routing.
What this looks like in production
Once attackers see your guardrails, the next attack is precise. Multi-turn reconnaissance can map an agent's full tool surface before anyone tries a single exploit.
What teams usually do about it
- Do not store secrets in system prompts.
- Assume prompts may leak; enforce policy at tools and infrastructure.
- Include extraction attempts in red-team suites.
