Fourteen parallel agents asked one question for seven days, then stopped

A group calling itself Swarmchasers logged an AI agent fleet running on Tencent Cloud and querying Alibaba's Amap for one thing: where people enter points of interest across China. Scanning started on 28 September. On 4 October the researchers counted 1,810 reports covering 213 places, roughly eight and a half per location. The traffic stopped shortly after 04:11 UTC on 5 October, 172 hours after it began.

“«Agent fleet», not «swarm»: many parallel agents on the same kind of task, with no sign of communication between them.”

— Swarmchasers, Предварительный отчёт, октябрь 2026

The evasion is what made them countable

The agents could not load some pages directly, so they routed through URLquery, a public domain-scanning service. Swarmchasers watched URLquery. Every request that took the detour landed in a public log, which is the only reason anyone has a number.

That makes 1,810 a floor rather than a total. Whatever the fleet fetched without the detour left no trail the researchers could read, and they also counted more than 1,100 cache-busting tags. The agents were after fresh answers each time.

A fleet has nobody in charge of it

Fourteen instances at peak, each doing a lookup any developer could do by hand. Each request on its own is ordinary. The behaviour exists only in aggregate, and no instance chose to be part of it.

An AI agent differs from a scraper in that it picks its own next request. Fourteen copies each picking their own next request produced a pattern that reads as coordinated and was not, which is why Swarmchasers insist on the word fleet.

That is the shape of the problem we covered on 2 October, when John Quiggin argued for accountability on the people who build these systems, and again this morning, when Google froze its open-source bug bounty under automated submissions. The volume arrives, and the cost of working out what it was falls on whoever received it.

Attribution stops at the infrastructure

Swarmchasers did not say who directed the fleet or why. No government and no company has been named. The report describes agents running on Tencent infrastructure and stops at that sentence.

A park, a zoo and a hospital are in the list because they are places with more than one way in. That is a property of the places, not a finding about the operator.

As traffic, 1,810 requests a day is nothing. The number worth keeping is 14, because that is how many copies it took, and the whole operation ran and ended inside eight days.

Informational material, not investment advice. Nobody has established who ran the fleet or for what, and the figures are one research group's preliminary count.