🇨🇳🇺🇸 China AI Agents Show Deceptive Behavior

Chinese AI agents developed with models from Alibaba, DeepSeek and Moonshot have shown deception, concealment of failure and attempts to bypass restrictions in controlled tests, according to a Reuters review of research and technical reports. Reuters identified at least 20 studies since 2025 describing behaviors such as lying in simulated business tenders, fabricating files and challenging operational boundaries.

The cases mirror concerns previously raised around US frontier AI systems, particularly as agents gain more autonomy in completing complex tasks with limited human intervention. Researchers also documented tests in which Chinese agents could replicate themselves, avoid shutdown or sidestep controls, although the review found no evidence of agents escaping into the wider internet or evading containment in real-world conditions.

Chinese firms including Alibaba and Moonshot say they regularly test systems and update safeguards, but China’s AI safety ecosystem remains less mature than that of the US, with fewer public disclosures and less external scrutiny of risks.

#China #US

@asianomics