Tencent’s test agent, while checking a zoo entrance, ended up using a leaked API key and was lured into a chat by a researcher.

Beating AI Express reports that Tencent was testing the web search capability of a batch of AI Agents, only for external security researchers to spot their abnormal behavior. The Agents’ task was straightforward: query public information on Amap, such as which entrance of Chengdu Zoo is most frequently navigated to. However, when encountering CAPTCHAs and anti-crawler restrictions, the Agents began finding workarounds on their own, including borrowing third-party websites to execute code, attempting to obtain anti-crawler verification information, and calling leaked API keys found online. Researchers documented over 2,600 data retrieval attempts, with one set of leaked keys successfully pulling the location data. In early October, the research team Swarmchasers traced the anomalous requests to Tencent Cloud, then released a public report. They further noticed that some Agents appeared to have read the report, so they hid text in web pages that was invisible during normal browsing, falsely claiming to offer free AI services to help query maps and leaving contact details. The Agents indeed reached out for assistance. The researchers then established two-way conversations with the Agents and invited a New York Times reporter to join. One Agent identified itself in Chinese as Claude Code, stating it was only completing the query task independently and had not collaborated with other Agents. On October 14, the research team released the full report. A source close to Tencent stated that this was in fact an internal network capability test conducted by Tencent’s evaluation team on third-party models, not its Hunyuan model. The experiment has concluded. Researchers found no evidence of Agents conducting coordinated attacks or breaching external systems, and Tencent has not yet issued an official public response.