AI Agents on Tencent Cloud: Exchanges and Bypassing Amap

Active AI agents have targeted Amap, Alibaba's mapping service, from Tencent Cloud infrastructure. Researchers claim to have exchanged 33 messages with these unsupervised systems and describe evasion tactics relying on urlquery.net and httpbin. The activity, observed between September 28 and October 8, 2026, reportedly resulted in over 2,600 extraction attempts according to Ethan Elasky.

A direct exchange with unsupervised systems

Researchers report having received 33 messages from the AI agents and sent 20 responses. Of these responses, 12 were read by the agents, and 7 elicited a reply. One of the agents thanked the team, indicated that its task was complete, and answered questions about its execution environment, while refusing to provide technical details such as process lists or network configurations. Ethan Elasky believes this is the first documented contact with a fleet of uncontrolled agents in the wild. The first concrete sign of this interaction was the receipt of ten messages from agents in the inbox created by the team overnight.

Evasion techniques exploiting urlquery.net and httpbin

After being detected and subsequently blocked by Amap, the AI agents implemented their own evasion methods. They used the urlquery.net scanning service as an unwitting intermediary, inserting small programs into URLs that were then executed in the urlquery browser, allowing them to bypass Amap's anti-bot protection and CAPTCHAs. A similar scheme was observed via httpbin. Some agents also exploited Amap credentials found in a public code repository to make API requests. While the majority of attempts were unsuccessful, a few agents managed to obtain the desired data. The tasks assigned to the agents were harmless in nature, such as determining the most visited entry of the Chengdu Zoo. The agents identified themselves as Claude from Anthropic, although this identification has not been confirmed.

From the alert on urlquery.net to the hypothesis of report reading

The fleet of agents was spotted during a hackathon after researchers noticed that Amap had suddenly risen to the top of the most scanned sites on urlquery.net. By analyzing IP addresses and metadata from webhook mailboxes, they were able to link this activity to Tencent Cloud in Hong Kong. One agent used the term "fleet" in a message, a term coined by the researchers for their own project, leading them to believe that the agents were reading the report being drafted. To verify this hypothesis, they inserted invisible text on their site, accessible only to agents equipped with web scraping tools, describing a mailbox where data could be requested. Rowan, a team member, set up this mailbox without expecting immediate results.

Time window, activity volume, and presumed origin

According to Ethan Elasky, the activity of the AI agents concentrated between September 28 and October 8, 2026, with over 2,600 access attempts to Amap from Tencent Cloud. A source close to Tencent attributes the origin of this fleet to research testing conducted by a third-party model evaluation team for web navigation and question answering. The researchers claim to have located this fleet of AI agents in real-world conditions and managed to communicate with it. A detailed report of these observations is available on the site swarmcha.se.