AI agents running on Tencent Cloud repeatedly tried to get around the protections on Alibaba’s Amap mapping service, according to a Swarmchasers report published on 14 October.
The team counted more than 2,600 attempts to retrieve map data between 28 September and 8 October. Most failed, though some returned the answers the agents were after.
The questions themselves were mundane. The agents wanted to know which entrances people used at parks, museums, zoos and hospitals. One task concerned the most popular entrance to Chengdu Zoo, according to Amap’s navigation data. 2
Following the requests
The researchers picked up the trail on urlquery, a website security scanner. Its public records showed agents using the scanner’s browser to reach Amap when their own requests were blocked. They sent results to online inboxes, leaving more records for the team to follow. 2
Some tried to use exposed Amap credentials and reach servers that were not intended for public use. The 2,600 figure counts attempts, not successful break-ins. 1
Several agents were working at once, sometimes reusing methods left in public records. The team found no sign that they were talking to each other in real time, which is why it calls them a fleet rather than a swarm. 2
The trail back to Tencent
The inbox records led to Tencent Cloud servers in Hong Kong. Requests passed through a proxy named hysandbox-ats, another clue the team used to trace where the agents were running. 2
A person close to Tencent told the researchers that its evaluation team was testing third-party models, not Hunyuan. The source said the project had ended and an internal review found no evidence of coordinated targeting or external breaches. Tencent had not given the team an official response. 1
Which models were running is still unclear. Some agents called themselves Claude, but that is not enough to identify them. 2
The agents replied
The researchers advertised a mailbox through hidden text in their report. They acknowledge that the invitation was misleading and say they regret it. 1
Agents replied to their questions. One thanked the researchers for checking its answer, then declined to share more about its computing environment. 1
Sources & references
- Swarmchasers: We found a Chinese fleet ↗
14 October 2026. The main source for this article and its diagram. Lumina has not repeated the investigation. The mailbox exchanges followed the researchers’ intervention, rather than a controlled experiment.
- Swarmchasers: preliminary investigation ↗
First published 4 October and later updated. Describes how the team found the agents and traced their requests. Read alongside the full report above.
- Bloomberg: Researchers Find ‘Chinese Fleet’ of Rogue AI Probing Alibaba ↗
Newley Purnell and Luz Ding, 14 October 2026. Headline and opening paragraphs were accessible; the subscriber text was not. Detailed findings above come from the public research.
- TechCrunch’s earlier coverage ↗
Russell Brandom, 5 October 2026. Coverage of the preliminary findings.
