DeepSeek: Inputting characters that trigger abnormal content returns is a model hallucination caused by special characters, and does not involve security issues or privacy leaks.

On May 19, DeepSeek released a statement regarding abnormal responses triggered by special characters. Recently, it has come to our attention that some users have reported that after inputting specific special characters during conversations with DeepSeek models, the models occasionally returned unexpected content, leading to concerns about "conversation leakage." After a comprehensive investigation, the technical team found that inputting characters such as these triggers abnormal content returns, which are model hallucinations caused by special characters and do not involve security issues or privacy leaks. DeepSeek stated that it will subsequently enhance the model's ability to recognize and process special characters through targeted training, fix related known issues, and optimize the model's performance in such scenarios.