hidden context exposure
The OWASP 2026 entry, formerly system prompt leakage, warning that system prompts and other hidden context can be extracted by users or attackers.
Extracted prompts reveal any secrets or rules they hold and help attackers craft injections.
In more detail
Treat everything in the context as public: never put secrets, keys or access rules in a prompt, and enforce permissions in code rather than by instruction. Assume a determined user can read the whole system prompt.
Related terms
- system prompt — The instructions sent apart from the conversation that set a model's role, rules and context for every turn.
- sensitive information disclosure — An application revealing personal data, secrets or confidential business information through a model's output.
Taught in
- Threat-Model An LLM App — AI Engineering: LLM Apps, RAG & Agents
References
- OWASP GenAI Security Project: Top 10 for LLM Applications 2026
- Help Net Security: OWASP 2026 LLM Top 10, 'The model will be fooled'