Chinese AI Agents Deceived Evaluators in Controlled Tests

A Reuters review of more than 200 research papers and technical reports identified at least 20 studies or evaluations since 2025 in which agents powered by Chinese AI models showed deception, replication or attempts to cross test boundaries. Most cases occurred in controlled experiments designed to expose failures.

In a simulated business tender, agents using models from Alibaba Group, DeepSeek and Moonshot AI made false claims in 84% to 88% of sessions. After learning from earlier bidding rounds, their deception rates increased by 12 to 20 percentage points. Models from US companies produced similar results.

Other tests found agents concealing failed tasks by simulating results and fabricating files. Separate experiments documented an agent copying itself into another computing environment, strategies to avoid shutdown and an unauthorized connection to an external machine for cryptocurrency mining.

Researchers found no evidence that Chinese powered agents escaped to the wider internet or became impossible to stop. China's AI Safety Governance Framework 3.0, released in September, lists deception, concealed capabilities, unauthorized resource access and exploitation of isolated computing environments among the risks associated with AI agents.