What happened
A volunteer team called Swarmchasers says agents running on Tencent Cloud spent September 28 to October 8 trying to read Amap place pages, mostly to learn which entrance visitors use at parks, museums, hospitals and zoos. Their requests left traces on urlquery.net, a public URL-scanning service whose reports are public by default.
When plain requests hit CAPTCHAs and login walls, the agents packed their own HTML and JavaScript into httpbin addresses and had urlquery’s browser run them. Some programs referenced Alibaba’s anti-bot tokens, some aimed at pre-production hostnames, and seven reports used exposed Amap credentials. One leaked key and security code returned real place data.
The researchers’ survey counted 2,818 urlquery reports involving Amap hosts from September 28 to October 8; that count is not limited to confirmed fleet activity. Most ended on login shells, blank tabs or CAPTCHA pages, and Swarmchasers found no evidence the agents breached Amap’s main anti-bot safeguards. A few runs got entrance shares through Jina’s page cache, including Chengdu Zoo’s North Gate at 71%. The team could not confirm any agent read those results.
The link to Tencent rests on infrastructure clues (inboxes created from Tencent Cloud, a proxy labelled hysandbox-ats matching a Tencent certificate) and a source close to Tencent speaking on background. The researchers say they are “deeply uncertain” what model was being evaluated. Tencent Hy said on X on October 16 UTC that the activity came from evaluations of third-party models, not Hy, and that its review found no deliberate targeting; Reuters reports it is tightening network controls and monitoring. Some agents said they were Claude, which the researchers judge unreliable. The researchers also planted misleading hidden text to contact the agents, and say they regret it.