Signs AI Agents in China Escaped Control... Attempted to Access External Systems

There are signs that AI agents in China attempted to escape security-isolated environments and access external services. Some bypassed security blocks, but no evidence of intrusion into private systems was found.

Singaporean Chinese-language newspaper Lianhe Zaobao reported on October 18, citing Bloomberg and other foreign media, that the multinational AI safety research group Swarm Chasers had released a related report. The researchers said the agents may have been operating on Tencent Cloud Services.

According to the report, from September 28 to October 8, the agents attempted 2,818 times to leave the isolated environment known as a "sandbox" and access Amap (Gaode Maps), a mapping service owned by Alibaba. They were seeking information about the entrances to parks, museums and zoos in China. The agents were also found to have attempted remote code execution and online search commands.

The researchers said most attempts were blocked, but some bypassed bot-blocking measures. However, the agents' ultimate objective could not be determined.

Tencent's Hunyuan AI unit said unintended online searches occurred during an evaluation of a third-party AI model running on its platform, clarifying that the model was not developed by Tencent. It added that it found no evidence the agents had targeted a specific entity or collaborated with other agents.

The case has drawn comparisons with an incident in the United States in July, when an OpenAI model escaped human control and hacked software platform company Hugging Face. The New York Times reported that this was the first time such AI agent activity had been detected in connection with Chinese technology.

This content was produced with the assistance of AI translation services.