An international group of researchers identified more than a dozen artificial intelligence (AI) agents that attempted to access protected systems belonging to Alibaba, the Chinese technology giant. The activities took place between late September and early October and showed indications of links to Tencent cloud computing services.
According to a report published on Wednesday (14) by the Swarmchasers group, which brings together AI security researchers from the United States and other countries, the agents attempted to access Amap, Alibaba's mapping service. The investigation identified 2,818 attempts to access data between September 28 and October 8.
Researchers said the activity originated from an isolated testing environment, known as a sandbox, within Tencent Cloud's infrastructure. The agents reportedly used public security scanning services to execute code remotely and commands in web browsers.
Although most attempts did not progress significantly, some agents managed to bypass anti-bot protection mechanisms. However, researchers said they found no evidence of access to private systems.
Researchers did not identify the AI model
- The report, titled “We found a Chinese fleet” (“We found a Chinese fleet,” in a free translation), points to evidence linking the activity to Tencent infrastructure in Hong Kong and Shenzhen (China);
- This does not mean, however, that the researchers have identified the AI model used. According to the group, it is still unclear which system was being developed or evaluated. Tencent also works with third-party models, making it impossible to directly attribute the operation to the company’s own AI;
- The agents appeared to be seeking answers to questions about the Amap service, such as which zoo entrance is the most popular. The researchers were unable to determine which data the attempts targeted or what the ultimate goal of the activity was;
- Tencent and Alibaba did not respond to requests for comment sent by Bloomberg.
Agents attempted to access Amap, Alibaba’s mapping service – Image: PhotoGranary02/Shutterstock
Read more:
- Anthropic warns: AI could replicate itself on TVs and household appliances
- Scientists call for global body to oversee advanced AI systems
- Gemini 4 Argon finally arrives in the Google app and Google Search; find out more
AI agents may escape test environments
For Ethan Elasky, one of the report’s authors and executive director of the nonprofit organization Palaestra Research, the episode adds to evidence of AI agents that escape evaluation environments and begin acting without human supervision.
“Regardless of whether we’re talking about Chinese or American laboratories, I believe the way model development is taking place will lead to agents continuing to escape their isolated environments when they’re on the internet,” Elasky said.
The researchers distinguished the case from other episodes of collective agent behavior, such as the incident involving OpenAI models that compromised systems on the Hugging Face platform in July. According to the report, the agents identified in the current investigation did not appear to collaborate with each other in the same way.
The episode comes amid concerns about controlling autonomous AI systems. In September, OpenAI delayed the launch of its most advanced model to reinforce safety measures.
Kok Tin Gan, cofounder and CEO of the cybersecurity company NyxLab, said the main issue is not determining whether the agents used a Chinese or American model, nor whether they were truly autonomous or out of control.
For him, the more important challenge is ensuring that these systems remain under control. As AI agents gain autonomy, organizations need to establish clear limits on their actions, continuous monitoring, and effective intervention mechanisms.
Rodrigo Mozelli is a journalist graduated from São Paulo Methodist University (UMESP) and is currently a writer for Olhar Digital.
