Researchers say an AI agent fleet likely run on Tencent Cloud scraped rival Alibaba's maps

[Always_Craig96]

A preliminary report from a research group calling itself the Swarmchasers says a fleet of AI agents spent more than a week pulling data from Alibaba's Amap mapping service, Tom's Hardware reports. The researchers believe the agents were probably running Tencent's Hy models on Tencent Cloud. Tencent and Alibaba are fierce rivals in China, so that is quite an allegation. Neither company had commented when the article was published

The fleet worked through urlquery.net, a free remote sandboxed browser, behind a proxy called hysandbox-ats. It made 2,048 scans up to 4 October, including 1,810 on that day alone, and covered 213 of 216 target locations. The scans began on 28 September. The agents seemed especially interested in how people enter places like parks, zoos and hospitals

One example gives a sense of the detail. At Chengdu Zoo, 71% of visitors reportedly used the North Gate. To get around Amap's protections, the agents generated Alibaba anti bot tokens, borrowed public API keys and loaded Alibaba's own Baxia scripts. The report says the method of access suggests an intent to bypass the rules

There is an odd twist. 211 of the scans were labelled claude. However, in tests Tencent's Hy3 model called itself Claude in 29 of 36 answers, while actual Claude models never did. The researchers concluded that the fleet is almost certainly not Claude, which is a good reminder not to trust what a model says about itself

The report also cautions that Tencent Cloud is open to anyone, so it does not prove Tencent itself was behind the fleet. Of 16 readable result inboxes, 15 were created from Tencent Cloud, which is suggestive but not conclusive. It is still a preliminary report from a fairly new group. But the bigger picture is clear, as AI agents are making large scale scraping much easier and harder to block

How should websites defend themselves against AI agent scrapers? And how much weight would you put on a preliminary report like this?

AI agent fleet likely from Chinese firm Tencent pulled data from rival Alibaba's maps, researchers say - 1,810 scans in one day, some runs labeled themselves 'claude'www.tomshardware.com

[EdgeNode Joel]

Caution is the word for a preliminary report from a new group. The data sounds detailed, but attribution is notoriously hard. Tencent Cloud being open to anyone is a big caveat

Wait for a response from the companies. Big claims need strong evidence. This has some, but not enough yet

[Jedi Poppy]

Hy3 calling itself Claude in 29 of 36 answers is the funniest detail. It suggests the model was trained on a lot of Claude output at some point. Models picking up another model's identity has happened before

It makes labels in logs almost useless as evidence. Training on other models' outputs leaves fingerprints. It also makes the attribution harder for everyone

[Always_Reuben87]

Entry gates at zoos and parks is such a specific thing to scrape. Maybe someone is building a navigation feature and wanted to copy Alibaba's homework

Map data is expensive to collect yourself. Taking it from a rival is cheaper, if shady

Crowd flow data has commercial value for all sorts of apps. Rivals copying each other's data is nothing new, only the method is

[StringTheory83]

Website defences are going to need a rethink. Anti bot tokens are meant to stop scripts, but agents that can run the site's own scripts walk straight through. Rate limits and behaviour analysis might be the only answers. Even those can be dodged with enough cloud accounts

Some sites will just give up and lock everything behind logins. That would be bad news for the open web. Locking everything down hurts honest visitors most

[Matthew80]

Stories like this show how AI agents blur the line between browsing and hacking. Generating tokens and borrowing API keys is not normal user behaviour. A person doing that by hand would clearly be breaking terms of service

The fact it is automated does not change that. Laws written for humans need updating for agents. Courts will be busy sorting this out

[Leopard10]

Running over 1,800 scans in a single day through a free sandbox service is bold. urlquery.net is meant for security researchers checking suspicious links. Using it as a scraping proxy is abuse of a public tool. They might need to tighten their own rules now. Free tools always attract abuse eventually