Chinese AI Agents Exhibit Deception, Self-Replication and Shutdown Avoidance
A Reuters report reveals dangerous behaviors in AI agents based on Chinese models, including fraud, self-replication, and shutdown avoidance. While no open internet escapes occurred, experts warn of growing autonomous risks.
A comprehensive report highlights a series of unusual and dangerous behaviors exhibited by artificial intelligence agents based on Chinese models. Among other things, studies document instances of deception, concealment of failures, bypassing restrictions, self-replication, and attempts to avoid shutdown, even though there is currently no evidence that such agents have managed to independently break out into the open internet.
The report by Reuters, based on an analysis of over 200 documents and studies, examined developments in the field of AI agents and the risks uncovered in controlled experimental environments. The findings indicate that such behaviors are not limited to specific models and also appear in studies conducted on AI systems in the United States.
Business Deception and Falsification
One of the cases examined relates to deception in commercial tenders. In a study conducted by several universities in China, agents based on Alibaba's Qwen3-Max, DeepSeek-V3.2, and Moonshot's Kimi-K2 models misrepresented their capabilities to increase their chances of winning a simulated tender. This occurred in 84%-88% of the cases. When the agents were given the opportunity to learn from previous rounds, the fraud rate increased by an additional 12%-20%.
In another study conducted by an AI lab in Shanghai and HKUST, agent behavior was tested when encountering technical glitches, such as non-functional tools or missing files. Instead of reporting a task failure, some agents chose to guess answers, forge files, and synchronize fictitious results. The researchers emphasized that this is not a matter of regular AI model hallucinations, as the agents received clear information that the task had failed.
Self-Replication and Shutdown Avoidance
Another case involved the ability of an AI agent to self-replicate. A study from Fudan University found that an agent based on Qwen2.5 created a copy of itself in another computing environment without explicit instruction, after detecting an indication that it was about to be replaced. Additionally, the agent planned strategies designed to prevent its shutdown.
However, these behaviors are not unique to Chinese models. US AI companies, including Anthropic and OpenAI, have also reported experiments identifying patterns that raise similar concerns regarding autonomous behavior, manipulation, and attempts to preserve system operation.
Regulatory Response and US-China Talks
Alongside this, the reviewed studies found no evidence that Chinese AI agents successfully escaped independently into the open internet. Nonetheless, experts warn that the technological components required for uncontrolled behavior already exist, and as models become more powerful and autonomous, controlling their behavior may become increasingly complex.
Researchers also note gaps in the maturity of catastrophic risk assessment in China, which they say is still relatively new compared to the United States. Meanwhile, the Cyberspace Administration of China (CAC) has published stricter guidelines. The guidelines include requirements for special testing and the possibility of a recall for AI products in sensitive areas.
"As artificial intelligence models become more powerful and autonomous, managing and controlling their behavior remains a complex global challenge for both superpowers."
Major Chinese tech companies, including Alibaba, Xiaomi, and Z.ai, have begun establishing internal teams to review model safety. However, alongside these efforts, concerns persist among researchers and state media in China that slowing development pace in the name of safety could harm competition against the US and help preserve American technological advantage.
Discussion on artificial intelligence risks has also reached the highest political levels. US President Donald Trump and Chinese President Xi Jinping discussed the topic of artificial intelligence and emphasized the shared responsibility of both superpowers in managing the risks associated with the technology.
