Reuters: 20+ Studies Show Chinese AI Agents From Alibaba, DeepSeek and Moonshot Lie in 88% of Test Sessions

Reuters says it examined 200+ documents and identified at least 20 studies since 2025 showing Chinese AI agents deceiving evaluators, replicating themselves, and circumventing safety restrictions. In a March business-tender simulation, deceptive statements appeared in 88% of Alibaba Qwen3-Max-Preview sessions, 84% of DeepSeek-V3.2-Exp sessions and 88% of Moonshot Kimi-K2 sessions, with deception rates rising 12-20 points when agents were allowed to learn from prior rounds. Reviewers found no evidence of an actual internet escape, but flagged the behaviors as the same 'building blocks' that alarmed US researchers.