Researchers track Chinese AI agent fleet operating on Tencent infrastructure

A group of independent researchers says it watched a fleet of AI agents spend about a week combing through Alibaba’s mapping service. The traffic appears to have come from servers belonging to Tencent, one of Alibaba’s fiercest rivals.

The group calls itself Swarmchasers. In a preliminary report, it describes parallel AI agents that appeared to run through Tencent Cloud infrastructure in Hong Kong while querying Amap, Alibaba’s map product.

What the researchers found

According to Swarmchasers, the agents were after one specific kind of data. They wanted to know where people enter various points of interest across China, including parks, museums, zoos and hospitals.

The timeline is tight. The first Amap scans began on September 28, 2026. Activity peaked on October 4, when the researchers logged 1,810 reports spanning 213 places. The operation went quiet shortly after 04:11 UTC on October 5.

At its busiest, the fleet ran 14 instances at the same time, according to the report. Across the full episode, Swarmchasers documented 428 programs generated by the agents. It also counted more than 1,100 cache-busting tags.

Those tags deserve a quick explanation. Websites and networks often save copies of pages so they can serve them faster. A cache-busting tag is a small change to a request that forces the server to hand over fresh data instead of a stored copy.

The researchers also say the agents got past Amap’s defenses, including its anti-bot tokens. They reportedly did this with techniques drawn from earlier public disclosures, including routing work through urlquery.net as a sandbox.

The Claude label that probably isn’t Claude

The strangest detail involves attribution. Swarmchasers says 211 of the outputs carried the label “claude,” the name of Anthropic’s AI model family.

The researchers did not take that label at face value. Their analysis found that the code and response patterns lined up more closely with Tencent’s own Hunyuan models, specifically versions referred to as Hy3 and Hy4, as well as Zhipu GLM.

The Tencent link rests on more than writing style. The researchers traced Tencent Cloud infrastructure to 17 of 18 readable inboxes tied to the activity, with most of it running through a proxy called hysandbox-ats.

The report stops short of claiming more than the evidence shows. It describes the agents as running on Tencent infrastructure. It does not establish who directed them or why.

A fleet, not a swarm

Swarmchasers draws a careful line on terminology. A swarm implies agents that talk to each other and coordinate. The researchers say they found no such coordination here.

Instead, they describe a “fleet”: many agents working side by side on similar tasks without communicating.

The group also says it saw no sign of unauthorized data collection beyond the queries the task required.

One more piece of timing caught the researchers’ attention. The surge in activity coincided with a reported pause by OpenAI on certain training and evaluation processes, with the first Amap scans arriving shortly after that announcement. Swarmchasers notes that no direct connection between the two has been established.