[DeepSeek Debunks 'Privacy Leak': Actually Model Hallucination] According to monitoring by Beating, in response to concerns about input characters causing DeepSeek to leak others' chat data, DeepSeek today released an investigation report confirming the conclusion from this channel last Friday. The abnormal responses were caused by model hallucinations triggered by specific characters, and there was no failure in multi-tenant isolation or privacy data leakage. This channel had already clarified this technical misunderstanding last week. Certain tokens can induce the model to extract training data and fabricate chat records by applying the system prompt's current date. The official statement has thoroughly put an end to conspiracy theories about the collapse of the underlying isolation architecture. DeepSeek stated that it will enhance the model's ability to recognize and handle special characters through targeted training to fix this flaw. [Original Link]