Prompts become logs, training disputes, and subpoenas—minimize sensitive data sent to vendors.
Rules
- Redact or tokenize PII before the model sees it
- Never put API keys, JWTs, or passwords in prompts
- Segment tenants—no cross-customer context in shared threads
- Honor deletion requests in your stored chat + index
Enterprise controls
Use VPC endpoints, zero-retention contracts, and regional residency when regulated. Document subprocessors.
Developer habit
Scrub logs in CI; use synthetic fixtures in eval sets. Rotate keys if leaked into a ticket pasted to ChatGPT.
Important interview questions and answers
- Q: Why tokenize PII?
A: Reduces exposure and may satisfy policy while keeping structure for the model.
Self-check
- List three data-minimization rules.
- What happens to prompts in vendor logs?
Tip: Run secret scanners on logs and tickets—developers paste keys into chats.
Interview prep
- Minimize data?
Prompts may be logged—redact PII and never send secrets.
- Tenant isolation?
Prevent cross-customer context leakage in shared threads.