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

CoinFeed reported on May 19th that DeepSeek issued a statement regarding character-triggered abnormal model responses. Recently, they noticed user feedback that after inputting certain special characters in conversations with the DeepSeek model, the model occasionally returned unexpected content, causing some users to worry about "conversation leaks." In response to this issue, the technical team, after a comprehensive investigation, found that the abnormal content returned due to inputting these special characters is a model illusion caused by the special characters themselves, and does not involve security issues or privacy leaks. DeepSeek stated that they 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.