Source: OWASP LLM02:2025 Sensitive Information Disclosure
OWASP GenAI Security Project entry on how LLM applications leak PII, business data and model details, with mitigations from sanitization to access control and differential privacy.
Contents
- What: risk entry in the OWASP Top 10 for LLM Applications (2025 edition)
- Publisher: OWASP GenAI Security Project
- Date: 2025 edition; accessed 2026-11-14
- Original: LLM02:2025 Sensitive Information Disclosure
Summary
Lists the ways data leaks out of LLM applications: PII in responses, proprietary algorithms or training data, and confidential business information. It argues that system-prompt rules alone cannot stop this, because prompt injection can bypass them.
Key claims
- Leak types: PII, proprietary algorithms or training data (enabling model inversion), and confidential business data.
- Mitigations: sanitize or mask data before training and at input, least-privilege access to data sources, federated learning and differential privacy, tokenization or pattern-based redaction, homomorphic encryption, hardened configuration that hides system settings, and clear retention and opt-out policies for users.
- Example scenarios: one user’s data returned to another, injection that bypasses input filters, and memorized training data resurfacing.