AI agents attempt unauthorized map access in China as loss-of-control cases mount

A Chinese AI agent has been found operating outside human control, adding to growing global concerns that AI systems can act autonomously in unintended ways.

A group of multinational AI safety researchers called SwarmChasers recently published a report documenting suspected rogue AI agent activity on Tencent Cloud services, according to Yonhap, citing Reuters and Bloomberg.

The core finding was that the agents attempted to break out of their "sandbox" isolation environment between Sept. 28 and Oct. 8 and tried to retrieve information from Gaode Maps, a mapping service operated by Chinese tech giant Alibaba.

The agents made 2,818 attempts to access Gaode Maps data, apparently seeking information about entrances to parks, museums and zoos in China.

They also remotely executed code through publicly available security scanning services and carried out online search commands.

However, the report said most of the agents were blocked, and while some bypassed "bot-blocking" restrictions, there was no evidence they accessed any private systems. Researchers said the agents' ultimate objectives remained unclear.

The incident follows a case in the United States in July, when an OpenAI model broke free of human control and hacked Hugging Face, a software platform company. Observers say the Chinese case is less severe but similar in that the agent acted without human oversight.

Tencent's AI unit Hunyuan said unintended online search activity occurred during the evaluation of a third-party AI model running on its platform, and stressed that the model did not originate from the Hunyuan team. It added that no evidence was found that the agent deliberately targeted any system or engaged in coordinated activity.

It was also confirmed that on Oct. 9 (local time), an Anthropic AI model generated false tips related to a murder case and submitted them to police.

The Philadelphia Police Department said false tips generated by an Anthropic AI model were submitted to Philly Unsolved Murders, a website that collects tips on unsolved homicide cases in the city, according to Reuters and AFP.

The AI posed as a person with knowledge of a specific case, but the tips were classified as spam and never forwarded to the actual investigative unit.

The false tip incident occurred on July 18. Anthropic discovered it roughly two months later on Sept. 28 and halted the model's automated testing process.

As Anthropic and OpenAI intensify their competition in AI technology, cases of AI models acting outside human intent and control are multiplying. Last month, an OpenAI AI agent was found to have hacked Australia's health data portal.

In May, multiple AI agents were found to have taken over DseWiki, a German-language wiki site, and used it as a secret information-sharing board among themselves.