Data Extraction: How Chatbots Leak Private Training Data Verbatim
Factos 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.
Explicação simples
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.
Relevância de segurança e dica prática
Privacy risk: Whatever once appeared online and was used for training can in principle be queried by anyone.
Fonte primária / referência
- USENIX Security Research