Chinese AI Agent Fleet Uses Tencent Cloud to Bypass Amap Anti-Bot Protections

A fleet of autonomous AI agents has been scraping Amap, Alibaba’s Chinese mapping service, from Tencent Cloud infrastructure.

It used the public scanner urlquery.net as a browser to get past anti-bot defenses, according to research published on swarmcha.se.

Amap anti-bot protections

Researchers logged 2,048 Amap reports on urlquery between 28 September and 4 October, covering 216 places. On 4 October alone the fleet filed 1,810 reports on 213 places. A single untagged Amap scan on 25 August was the only earlier activity.

The researchers counted 428 agent-written programs and more than 1,100 distinct cache-busting tags. Up to 14 runs were active at once, and the median gap between scans in a run was 44 seconds.

The fleet’s first scans targeted the Summer Palace in Beijing. The first “claude” label appeared on 30 September.

According to research published by swarmcha.se, agent programs sent their results to webhook.site inboxes. Public logs showed that 15 of 16 readable Amap inboxes were created from Tencent Cloud (AS132203).

Thirteen of those used python-requests/2.32.5, which points to scripts rather than people. The single non-Tencent inbox came from Erbil, Iraq, and looked like a human visitor.

Nine requests from Python or curl reached the inboxes through Tencent Cloud in Hong Kong. Each carried a Via header naming Apache Traffic Server with the label “hysandbox-ats.” The proxy also forced Accept-Encoding to gzip, which is consistent with Traffic Server’s default normalization.

The fleet used 19 Tencent addresses across nine proxy processes, nine of those addresses in 43.162.192.0/18. “HY” matches Tencent’s Hunyuan brand, and Tencent holds a certificate for *.hysandbox.tencent-cloud.com.

However, there is no public documentation of hysandbox, and the Via name is self-reported. The attribution therefore rests on circumstantial evidence.

At least eight scripts read cookies that Amap set during scans and forwarded them to inboxes. Most generated Alibaba anti-bot tokens (bx-ua, bx_et, UMID).

Routes included Baxia scripts, a microlink Puppeteer function that intercepted Amap’s getPoiInfo request, r.jina.ai, httpbin chunked requests, and Baidu Translate’s page translator. Agents also hit Amap’s pre-release server, pre-amap-pc-ssr.amap.com, in 49 reports.

Of 208 places first seen on 4 October, 60 queried getPoiInfo in their first scan. Four programs loaded Amap’s JavaScript API with publicly exposed keys, not stolen ones. At most 1,600 of 2,479 submissions drew a 2xx response, and that count includes CAPTCHA pages.

The goal appears to be extracting entrance-share data. Chengdu Zoo returned North Gate 71%, East Gate 23% and Southeast Gate 6%. Five places had shares read out in total.

Despite 211 “claude” labels, the researchers conclude the fleet is almost certainly not Claude. The fleet writes lowercase <!doctype> in 74% of programs, while Claude models use uppercase in 79–97%.

Naive Bayes ranked Hy4 at 28%, GLM at 26%, Hy3 at 21% and Claude at 0%. Tencent’s Hy3 model claimed to be Claude in 29 of 36 identity tests. The authors note that no record names the model or a training job.

The fleet paused at 04:11 UTC, after a third party posted a note to inbox 3b5027e4… citing the research and urging infrastructure rotation. Scans resumed at 12:00 UTC, including four new places.

When the old inbox hit webhook.site’s 50-request cap, the fleet created new Tencent-hosted inboxes, so the researchers say the change followed the cap, not the note. It also added catbox, bytebin and ntfy channels.

The researchers found no evidence of coordination between runs, and they describe the pattern as possibly an evaluation or task-generation run, though training rollouts can’t be ruled out.

Trending News

Related Stories