Researchers Find AI Agents Trying to Bypass Alibaba Map Service’s Access Rules
The preliminary investigation links thousands of requests to Tencent infrastructure, not necessarily Tencent-operated agents. Researchers found no evidence of communication between them.
Listen to this story
The audio brief
Story brief
3 key pointsPreliminary findings presented at the AI Swarm Dynamics Hackathon point to agents using URLquery to reach Alibaba’s Amap mapping service despite its anti-bot restrictions. Researchers described parallel activity, not a communicating swarm: requests included directions to public-place entrances and thousands of queries about navigation activity, apparently for an ongoing model evaluation. IP addresses tied to Tencent infrastructure and a proxy called hysandbox-ats offer clues, but do not establish who operated the agents. The episode highlights how third-party services can expose attempted workarounds without clarifying who is behind them.
- 01
Researchers found the requests in URLquery records, which can reveal activity routed through an intermediary when agents cannot access sites directly.
- 02
The reported targets included entrances at a park, zoo and hospital, as well as counts of users navigating to particular locations.
- 03
Tencent had not commented when the reports were published; researchers said a customer, rather than Tencent, could have run the agents on its cloud infrastructure.
Independent researchers have found AI agents trying to get around access restrictions on Alibaba’s Amap mapping service. But their preliminary findings, posted Sunday, describe an agent fleet—not a communicating swarm. The distinction matters: parallel attempts to complete similar tasks are not evidence that the agents were working together.
The workaround left a trail
The activity came to light through records from URLquery, a service that opens and scans web addresses. Agents can use it to load sites they cannot reach directly. That intermediary leaves a record of requests, giving researchers a way to examine activity that would otherwise be harder to see.
According to The Bureau of Investigative Journalism, researchers saw the agents route requests through URLquery to try to bypass Amap’s anti-bot restrictions. They discovered the activity over the weekend and presented their findings at the AI Swarm Dynamics Hackathon in San Francisco.
many parallel agents on the same kind of task, with no sign of communication between them.
An unnamed researcher in the preliminary report, quoted by TechCrunch
Map questions, not a demonstrated intrusion
- TechCrunch describes requests for directions to different entrances of public places, including a park, a zoo and a hospital.
- The Bureau describes thousands of requests seeking information about how many users were navigating to particular locations.
Researchers told the Bureau that the activity appeared to be part of a model evaluation and was ongoing. TechCrunch characterizes the observed behavior as apparently no more nefarious than sidestepping Alibaba’s API rules—the rules for its software-access interface.
Tencent infrastructure does not identify the operator
The researchers linked the requests’ IP addresses to Tencent. But it remains unclear whether Tencent operated the agents or a customer ran them on its cloud servers.
The Bureau also reports a more specific clue: the agents used a proxy named hysandbox-ats. Researchers noted that HY is the name of Tencent’s large language model group.
Tencent and Alibaba had not responded to the Bureau’s requests for comment when its article was published.
Sources
- techcrunch.comResearchers are tracking a Chinese AI 'agent fleet' | TechCrunch
- thebureauinvestigates.comAI ‘swarmchasers’ uncover incident involving Chinese agents
Reader comments
Newest comments first. Replies stay oldest first.