Data Extraction: How Chatbots Leak Private Training Data Verbatim
Hechos validados
Carlini et al. (Google DeepMind, USENIX Security). Through targeted prompts, researchers made GPT models disclose original text fragments from training data (including personal names, emails, and phone numbers) verbatim.
Explicación sencilla
Elephant memory: A chatbot does not only summarize — sometimes it has memorized whole book pages or phone numbers and blurts them out when asked the right questions.
Relevancia de seguridad y consejo práctico
Privacy risk: Whatever once appeared online and was used for training can in principle be queried by anyone.
Fuente primaria / referencia
- USENIX Security Research