AI TRENDS | DeepSeek Addresses Model Response Anomalies Triggered by Special Characters

On May 19, DeepSeek released a statement regarding anomalies in model responses triggered by special characters. According to Jin10, users recently reported unexpected content being returned during interactions with the DeepSeek model after inputting certain special characters, raising concerns about potential 'conversation leaks.' After a comprehensive investigation, the technical team identified that the anomalies were caused by model hallucinations triggered by these special characters, and confirmed that there are no security or privacy issues involved. DeepSeek announced plans to enhance the model's ability to recognize and handle special characters through targeted training, aiming to resolve known issues and improve performance in such scenarios.