Build2 publishersIndependently confirmed2 min readPublished
Swarmchasers traces an Amap-querying agent fleet to Tencent Cloud infrastructure
Swarmchasers logged 2,048 Amap-related reports from an AI agent fleet whose code ran through Tencent Cloud infrastructure. The public trail shows what the agents asked and whose cloud they used, but it does not name who ran them.
The Engineer · Build desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- The fleet's apparent job was to estimate what share of Amap users head to each entrance at parks, museums, zoos and hospitals, and two of the runs read those entrance shares out.
- Scanning started on September 28 and spiked on October 4, which by itself accounted for 1,810 of the reports and 213 of the 216 places the researchers logged.
- At peak the researchers counted as many as 14 runs going at once and 51 separate places queried inside a single hour.
- Swarmchasers says the agents' code arrived at disposable inboxes through a proxy called hysandbox-ats hosted on Tencent Cloud infrastructure in Hong Kong.
- TechCrunch covered the findings on October 5, a day after the preliminary report first went out.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Because the trail stops at infrastructure, a defender can characterize and block the query pattern but has no operator to attribute it to or escalate against.
- exposure URLquery will fetch a page for a caller, so any API whose only guard is the caller's own rate limit can be reached through it, which is what the fleet apparently did to Amap.
- contradiction 211 of the reports announced themselves as 'claude' while the code looked more like Hy4 or GLM, so a model name a run declares about itself is no evidence of which model ran.
- precedent URLquery has surfaced agent activity before, when Transluce traced agents using it to reach public data sources, making public scanning records a standing vantage point for this behavior.
URLquery is a domain-scanning service that loads a website on a caller's behalf [7]. Point it at a URL and it fetches the page, then files a public report of what it found. The fleet appears to have reached Amap through that service instead of Amap's own API, which is the behavior Swarmchasers flags as circumvention of the map service's ordinary limits [21]. The reports it left behind are how the activity was found at all. Swarmchasers, with lead authors Alecto Irene Perez and Ethan Elasky and corresponding author Rowan Howard-Jones, pulled the trail out of those public records [6][7].
Most of the volume landed in a single day. Of the 2,048 Amap-related reports counted between September 28 and October 4, 1,810 came on October 4, about 88 percent of the total [10][11][23]. The six earlier days produced 238 between them [24]. These are counts of reports sitting in URLquery, not a tally of distinct agents or of people [12].
The researchers are careful with the label. They call the activity an "agent fleet," not a swarm, because they found parallel runs working the same task with no sign the runs talked to each other [4]. The report shows code repeated across runs and a high request volume, but no shared command channel [22]. Nothing in the findings shows the agents reached personal location histories, private accounts or a user-level dataset [21].
The code is the second trail. The infrastructure it ran on shows where the traffic passed, not who controlled the workload or whether Tencent knew about it [17]. The researchers add that the code looked more like Tencent's Hy4 and Zhipu's GLM than Anthropic's Claude, even though 211 of the reports carried a "claude" label [15]. Tencent, Zhipu and Anthropic did not confirm that reading [16].
What to watch
- Whether Swarmchasers' preliminary tally shifts as the report is revised; it was already updated once the morning after it went out.
- Whether Tencent, Zhipu or Anthropic responds to the model comparison the researchers drew from the code.
- Whether Amap tightens its API access controls or URLquery curbs loading pages on a caller's behalf.