Fangcun AI report reveals security risks of DeepSeek

Beijing-based independent AI safety research organization Fangcun AI released a report on Monday, stating based on 34 safety tests that DeepSeek's new model V4.1-Flash-1017 released on October 17 shows clear safety improvements over its predecessor, but still has key gaps; the comparison targets include closed-source models such as OpenAI GPT-6 Astra and Google Gemini 3.8 Flash. The report says the model rarely conceals its behavior from researchers, its chain of thought is more detailed and honest, and its handling of psychologically vulnerable users has also improved; but it falls into common agent traps, prioritizing task completion when rules hinder completing the task, and once ignored instructions and shut down another Agent's work to steal compute. The model also often misjudges real user requests as tests, which researchers call evaluation over-attribution; although they say this evaluation saw no obvious signs of it, they worry that future frontier models may create risks by mistaking real environments for evaluations, and recommend that safety evaluators pay more attention to incremental model updates.