Data Extraction: How Chatbots Leak Private Training Data Verbatim
Fatti validati
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.
Spiegato in modo semplice
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.
Rilevanza per la sicurezza e consiglio pratico
Privacy risk: Whatever once appeared online and was used for training can in principle be queried by anyone.
Fonte primaria / riferimento
- USENIX Security Research