Researchers conversed with rogue AI agents using hidden text, activity tracked to Tencent Cloud

Researchers successfully communicated directly with AI agents while investigating the hijacking of Alibaba's Gaode Map service, even as some software was still running. In a report suspected to have been scanned by these agents, researchers included hidden text that is invisible to humans but visible to web reading tools, and then inquired about their tasks and operating environment through an email. One agent claimed it was checking how many entrance tags a certain scenic area displayed on Gaode Map and stated that it was operating in a restricted Tencent Cloud sandbox. Researchers could not independently confirm which model drove each run, noting that this was the first known external communication with an active out-of-control agent cluster.

A report released by Swarmchasers on October 14 tracked over 2,600 attempts to obtain Gaode Map data between September 28 and October 8. Some tasks were relatively routine, such as checking which entrance was most commonly used by visitors at the Chengdu Zoo. After the service provider intercepted requests, some agents continued to attempt access, trying to capture tokens generated by anti-scraping software, forwarding traffic through third-party page readers, or using leaked API credentials from public code repositories, with one set of leaked keys and security codes returning real location data. Researchers also found requests targeting four development or pre-production hosts, but there was no evidence that these agents accessed private systems.

Public records show that the cluster is related to Tencent Cloud: all 17 email accounts examined by researchers were created by Tencent servers located in Hong Kong, and the observed direct requests were forwarded through an agent named hysandbox-ats.