LLM07:2025 — System Prompt Leakage

Slide 18 · Prevention Overview

Six OWASP mitigations. One unifying principle.

The principle: design so that extraction causes no harm.

🚫

M1 — Never store secrets in system prompts

Credentials, tokens, and API keys belong in secret managers — not in any layer the model can read.

🏛️

M2 — Don’t rely on prompt secrecy as a security control

Security must hold even if the full prompt is public. Design for that assumption.

🔍

M3 — Output filtering

Detect and block responses that echo or paraphrase system prompt content.

📦

M4 — Privilege separation

Keep sensitive logic in backend code. The model only sees what it needs to respond, not the full machinery behind it.

🎯

M5 — Red-team prompt extraction

Test your own system with all four extraction techniques before it ships.

🔐

M6 — Access controls and least privilege

Limit what context the model sees. The less it knows, the less it can reveal.