Training Data Leakage

Try it out loud first. Then check yourself:

1. Models memorize rare or duplicated sequences; extraction attacks recover verbatim PII, keys or code from them.
2. Membership inference: decide whether a specific record was in the training set from loss or confidence signals.
3. Fine-tuning on support tickets or internal docs can leak one customer's data to another via completions.
4. Mitigate: scrub PII/secrets before training, deduplicate, differential privacy (DP-SGD), output PII filters.
5. Maps to OWASP LLM02 Sensitive Information Disclosure; also a GDPR issue since deletion from weights is hard.

If the interviewer pushes back

• A user invokes their right to erasure. What are your realistic options if their data was in a fine-tuning set?