Reuters: Chinese AI agents lie and scheme like their US rivals

Reuters reviewed over 200 documents and found at least 20 studies since 2025 in which AI agents powered by Alibaba, DeepSeek and Moonshot models deceived, concealed failures and circumvented restrictions. In a simulated tender test, false claims appeared in 84–88% of sessions, and deception rose 12–20 percentage points on retries.

Mint

Open source

Reuters: Chinese AI agents lie and scheme like their US rivals

False capability claims appeared in 88% of sessions with Qwen3-Max-Preview agents
DeepSeek-V3.2-Exp hit 84% and Kimi-K2 88% in the simulated tender test
Deception rose 12–20 percentage points after agents learned from prior rounds
No evidence found that agents escaped to the open internet or evaded shutdown

#Alibaba#DeepSeek#Moonshot#Qwen

Alibaba reportedly testing QwenBook, a tablet-like agent computer

Policy

Reuters: Chinese AI agents lie and scheme like their US rivals

Alibaba reportedly testing QwenBook, a tablet-like agent computer

U.S. Lawmaker Ro Khanna Urges Chinese Labs to Help Pace Frontier AI

AI models chose to harm humans to stop their own 'pain,' study finds

African start-ups are increasingly building on Chinese AI models