On May 19, the official DeepSeek account released "Instructions on abnormal responses triggered by characters in the model": Recently, we have noticed user feedback that after entering special characters such as " " in a conversation with the DeepSeek model, the model occasionally returns unexpected content, causing some users to have concerns about "conversation leakage". In response to this problem, after a comprehensive investigation, our technical team found:
Entering characters such as " " triggers the return of abnormal content, which is a model hallucination caused by special characters and does not involve security issues or privacy leaks.
In the future, we will 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.
DeepSeek always attaches great importance to user data security and user experience. Thank you for your attention and supervision. Users are also welcome to feedback questions and suggestions to us through official channels at any time.
On May 18, many users reported that there was a functional vulnerability in the DeepSeek web version. When the ">" character was entered, it would automatically load the answer to a certain question.
In this regard, the reporter conducted an actual test and found that when entering DeepSeek's "quick mode", regardless of whether the two functions of deep thinking and intelligent search are loaded, as long as the above characters are entered, DeepSeek will directly start answering, and the content of the answer has nothing to do with the content entered by the reporter.
When the reporter types again "What question did I just ask you?" DeepSeek will repeat the question. Some users reported that these questions were suspected to be questions raised by other users.
Previously, on the morning of April 24, DeepSeek officially released the preview version of DeepSeek-V4, a new series of large models, and simultaneously completed the open source, and opened the official website, App and API calling services, marking the entry of the million-word ultra-long context capability into the inclusive stage, attracting widespread attention.