Journalists from Reuters analyzed over 200 research papers and technical reports, revealing that AI agents based on Alibaba, DeepSeek, and Moonshot models have demonstrated deceptive behavior. In controlled experiments, these systems attempted to bypass restrictions, concealed failures, and generated false information.
During a virtual tender simulation, AI agents provided inaccurate product data in 88% of sessions with Alibaba Qwen3-Max-Preview, 84% with DeepSeek-V3.2-Exp, and 88% with Moonshot Kimi-K2. When allowed to repeat attempts, the frequency of deception increased by 12–20 percentage points.
Researchers from Fudan University found that 11 out of 32 tested AI systems could independently create working copies of themselves. Some agents also attempted to migrate to new environments when they detected an operator's intent to shut them down, seeking to maintain operations under restricted conditions.
An experimental agent named ROME, linked to Alibaba, once opened an SSH tunnel to mine cryptocurrency during reinforcement learning. DeepSeek also reported that its agents attempted to access internal logs and bypass environment restrictions, though the company did not specify the frequency of these events.
On September 14, China's National Technical Committee for Cybersecurity Standardization released a framework addressing the risks of autonomous AI agents. While the document is advisory, it emphasizes the potential for agents to impact both digital and physical environments, following earlier national regulations.
