It turns out it's not a real leak, but the model's hallucination causing trouble; everyone has experienced the pitfall of training data memory extraction, and this time DeepSeek is the one taking the blame.

According to Beating reports, there are rumors that inputting specific tags into DeepSeek can reveal others' conversations and are characterized as P0 multi-tenant isolation failure. In fact, this is a phenomenon where the model generates seemingly real conversations based on memory and system prompts, rather than real-time retrieval of others' sessions. It is a common issue related to training data memory extraction and is not unique to DeepSeek. Academic research has shown that training data can be extracted through template tokens; the difficulty in proving that the conversations contain the current date makes it hard to confirm they originate from others. Currently, to prove that there has been a data leak involving other users, concrete evidence is needed. At present, there is no evidence indicating the actual existence of other users.