How should you reduce the risk of sensitive data reaching logs and prompts?
Why D is the answer
Send only the fields required, and redact before data enters the system
Data minimization is the strongest control, because what never enters the system cannot leak. Prompts cannot guarantee it.
In detail
- Data minimization is the strongest control.
- What never enters the system cannot leak.
- Check logs as well as prompts, because data leaks into both.
Use case
An illustrative scenario, not a real incident.
A claims tool sends whole records to the model and logs them. Sending only the fields needed, with identifiers redacted first, removes most of the risk.
See it in different ways
flowchart LR A["Classify the data"] --> B["Apply regulation"] B --> C["Choose deployment and vendors"] C --> D["Minimize and redact"] D --> E["Log and audit"]
mindmap
root((Compliance))
Health data
Contract terms
EU personal data
Lawful basis
Where it is processed
US federal
FedRAMP
All
Data minimization
Audit trailflowchart LR S0["Source record"] --> S1["Keep required fields"] --> S2["Redact identifiers"] --> S3["Send to the model"]
Key point: Minimize data before it enters the system.
Why the other options are wrong
- A. Delete all logs every day, so that nothing sensitive remains in storage for very long
- The data is still exposed until the logs are deleted.
- B. Encrypt the prompts after the model has finished processing them
- By then, the model has already processed it in the clear.
- C. Ask the model never to repeat anything that looks like sensitive data
- A prompt cannot guarantee this, and the data has already entered.