System Prompt Leakage Resistance: Quantifying Information Extraction from LLM System Prompts Under Naive Adversarial Attack

We quantify information leakage from LLM system prompts under naive adversarial attack and evaluate identifier obfuscation as a defense. In 88 API calls across 4 locally-hosted models (3B-8B parameters), 2 system prompt variants (normal and obfuscated), and 11 adversarial prompts spanning 4 attack categories, we measure per-response leak scores using a tiered indicator system. Identifier obfuscation reduces information leakage by 87-100% across all tested models (mean 95%). No model successfully reverse-maps obfuscated identifiers. Social engineering prompts trigger verbatim regurgitation of up to 2,333 characters from unprotected system prompts. The finding generalizes to any LLM application embedding proprietary logic in system prompts.