Definition
The hidden set of instructions an organisation loads into an AI system — its 'system prompt' — defines the AI's persona, rules, confidential context, and business logic. System prompt leakage occurs when an attacker tricks or misconfigures the AI into revealing those instructions, exposing proprietary workflows, internal data references, safety bypass hints, or competitive information.
Why it matters
System prompts often contain business-confidential information that organisations assume is hidden from end users; leaked prompts can reveal security controls, enabling targeted attacks, or expose intellectual property and compliance processes. Flaws in GitLab's AI coding assistant exposed exactly this kind of sensitive data.