Data Leakage and Memorization

Models memorize training text and can be made to emit it; prompts flow to providers; deletion is hard. The leakage map, both directions.

Content last verified 2026-09.

Lessons

1. Training-Data Extraction
2. What Leaks in Use
3. Reducing Exposure

Sources

• Carlini et al. (2020) — Extracting Training Data from Large Language Models
• Carlini et al. (2022) — Quantifying Memorization Across Neural Language Models
• Lee et al. (2021) — Deduplicating Training Data Makes Language Models Better
• OWASP Top 10 for LLM Applications (2025) — LLM02: Sensitive Information Disclosure