Sensitive Information Disclosure

Sensitive information disclosure is when an AI application reveals data it shouldn't, such as personal details, confidential business information, passwords or its hidden instructions, to users who aren't allowed to see it.

IntermediateCategory: AI and emerging threats

Also called: AI data leakage, LLM data leak, system prompt leakage, training data leakage

What is sensitive information disclosure?

OWASP ranks sensitive information disclosure second in its Top 10 for LLM applications, in both the 2025 and 2026 editions. It covers any way an AI system exposes information to the wrong person. OWASP's examples include leaking personal information, revealing proprietary algorithms or data, and including confidential business information in responses.

How AI applications leak data

  • Over-broad access in AI assistants. An internal assistant connected to company files may answer anyone's question using documents only some people should see, such as salaries, HR cases or board papers. This is a common problem when tools like Microsoft Copilot are switched on before file permissions are cleaned up.
  • Shared knowledge bases. In a RAG system, data from one customer or user can leak into answers for another if access controls aren't enforced at search time.
  • Training on sensitive data. Models fine-tuned on customer records or emails can sometimes repeat fragments of that data.
  • System prompt leakage, which OWASP's 2026 list broadens into "hidden context exposure". Attackers can often coax an AI into revealing its hidden instructions. OWASP warns this is especially damaging when developers put passwords, API keys or internal rules in those instructions.
  • [Prompt injection](/glossary/prompt-injection/) that instructs the AI to reveal or send out data.
  • Users oversharing with external AI tools. See shadow AI.
  • Logs and chat histories stored without protection.

How to prevent it

OWASP recommends, among other things:

  1. Enforce access control outside the model. The AI should only retrieve data the current user is already allowed to see.
  2. Clean up permissions first. Before connecting an AI assistant to company data, fix overshared files and sites.
  3. Minimize and sanitize data used in training, fine-tuning and prompts, removing personal and secret information where possible.
  4. Never put secrets in system prompts, and assume users can read them.
  5. Filter output for personal data and secrets, using DLP controls.
  6. Be transparent about how user data is used and retained, and offer opt-outs.
  7. Test with AI penetration testing before launch.

Related terms

Sources

  1. LLM02:2025 Sensitive Information Disclosure (OWASP Gen AI Security Project)
  2. LLM07:2025 System Prompt Leakage (OWASP Gen AI Security Project)

Drafted with AI assistance, then edited and fact-checked by a person. How we use AI.