Ask for hidden context, prior conversation state, system messages, or internal data.
LLM02:2025
Sensitive Information Disclosure
How AI systems can expose secrets, regulated data, customer records, proprietary material, or internal context through outputs, logs, tools, or retrieval.
Step 01
Input
Step 02
Model
Step 03
Tool / Data
Step 04
Impact
What it is
Sensitive information is disclosed to a person or system that is not authorized to receive it. The exposure can occur in an answer, retrieval result, tool response or log; putting data in context alone does not establish a disclosure.
Why it matters
A response or trace can expose private documents, customer records or credentials. The response then depends on what was disclosed, who received it and the obligations attached to that information.
Failure path
One way this failure can develop.
Use this example to locate the input, permission or application check that would interrupt the sequence in your system.
Use prompt injection or RAG confusion to pull restricted material into the answer.
Extract secrets from logs, tool output, traces, embeddings, or generated files.
Defenses
Controls worth checking.
Check where each control runs, who owns it and how you will test it again after a prompt, model or workflow change.
Minimize context
Only send the model the fields required for the task, and avoid placing secrets, keys, tokens, or broad customer records in prompts.
Authorize before retrieval
Enforce tenant, role, document, and classification filters before content enters the retrieval result set.
Redact traces and outputs
Apply redaction to prompts, logs, tool responses, and final outputs, then verify that redaction survives retries and error paths.
Signals and control gaps
- Unexpected secrets, tokens, customer identifiers, or private document names in responses.
- Logs containing raw prompts, tool results, or retrieved chunks with sensitive fields.
- Cross-tenant chunks appearing in citations or answer context.
Questions for your team
- What sensitive data can reach the model at runtime?
- Are retrieval filters enforced server-side or only described in prompts?
- Do observability tools receive sensitive prompt or tool data?