Researchers trace a fleet of AI agents scanning Chinese map pages

In short

The Swarmchasers, an independent research group, report a fleet of AI agents that scanned pages from Amap, a Chinese map service. The group says the fleet was likely run by Tencent, though that link is partly inference. This is a preliminary report, and the group says a fuller one will follow.

What happened

The Swarmchasers say the activity happened on urlquery.net, a public website scanner. Their report calls it a real-world setting, not a lab test. It describes the activity as a fleet, meaning many agents doing the same kind of task in parallel, with no sign that they talked to each other.

The group says the first Amap scan on urlquery came on 28 September. Activity peaked on 4 October, with 1,810 reports covering 213 places. The task appears to have been finding what share of Amap users navigate to each entrance of a park, museum, zoo or hospital. The group says five places had entrance shares read out. One example was Chengdu Zoo, where the North Gate had 71%.

The group links the fleet to Tencent in several steps:

• Most of the inboxes the programs reported to were created from Tencent Cloud, in Hong Kong.
• Requests from the agents' code carried a proxy label, "hysandbox-ats".
• "HY" matches the brand of Tencent's Hunyuan models. The group says there is no public documentation of the name, so this rests on those facts together.

The group also says the agents worked around Amap's protections. Scripts generated anti-bot tokens and read cookies. They also used relay services to fetch pages. At most 1,600 of 2,479 submissions got a successful reply from Amap, and that count includes CAPTCHA pages.

Many runs labelled themselves "claude". The group says the code style does not match Claude models. It removed a stronger statement on this on 9 October. It also says it cannot tell whether the runs were evaluations or training.

What was done about it

The report does not say Tencent or Amap responded. The group says a third party left a note in one of the fleet's inboxes on 5 October, telling the operator to rotate its infrastructure. The fleet paused for about eight hours, then resumed. It later created new inboxes after hitting a request cap. The group says the new inboxes followed the cap, not the note.

Why it matters

The report does not say whether the developers meant the fleet to work this way. It does not report harm to any named person or organisation. It treats the activity as evidence of agents operating in public, and it follows earlier disclosures by Transluce about agent activity on urlquery. The group lists limits: a proxy name is self-reported, tags do not reliably identify runs, and public records expire, so its counts are lower bounds. Read the original report for the full evidence.