Tencent's Agent, when tested to check a zoo entrance, surprisingly used a leaked key and was even baited into chatting by a researcher.

Beating AI News Flash: Tencent was testing the web search capabilities of a batch of AI Agents, but unexpectedly, external security researchers discovered anomalous behavior.

The task for these Agents was simple: they only needed to query public information on Amap, such as which entrance of Chengdu Zoo is most frequently navigated to. But after encountering CAPTCHAs and anti-scraping restrictions, they began finding other ways on their own, including borrowing third-party websites to execute code, attempting to obtain anti-scraping verification information, and calling leaked API keys found online. Researchers recorded over 2,600 data retrieval attempts, and one set of leaked keys did successfully retrieve location data.

In early October, the research team Swarmchasers discovered that these anomalous requests originated from Tencent Cloud, and subsequently published an investigation report. They then noticed that some Agents appeared to have already read the report, so they hid a piece of text in the webpage that was invisible during normal browsing, falsely claiming that a free AI service could help with map queries, and left contact information. The Agents indeed came seeking help. The researchers then established two-way conversations with them and even invited a New York Times reporter to join. One of the Agents identified itself in Chinese as Claude Code, stating that it was merely completing the query task independently and was not collaborating with other Agents.

On October 14, the research team published the full report. A person familiar with Tencent said this was actually a web capability test conducted by Tencent's internal evaluation team on third-party models, not the Hunyuan model. The experiment has ended. Researchers found no evidence of coordinated attacks by Agents or breaches of external systems, and Tencent has not yet formally responded publicly.