Chinese AI Agent Fleet Caught Targeting Alibaba Amap From Tencent Servers

TL;DR

• Independent researchers tracked thousands of automated agents running on Tencent Cloud IPs systematically probing Alibaba's Amap mapping APIs in late September.
• The fleet mimics human map behavior at machine scale — querying routes, traffic, and POIs in patterns consistent with training autonomous navigation and location-aware AI agents.
• The incident highlights a new front in China's AI arms race, where Tencent, Alibaba, and ByteDance are racing to build real-world-capable agents that need massive live geospatial data.

A Strange Spike in the Logs

It started, as many internet mysteries do, with an anomaly no one was supposed to see. In mid-to-late September, independent security researchers monitoring public API traffic noticed an unusual surge in requests hitting Alibaba's Amap, China's dominant mapping service.

The volume alone was staggering — millions of route-planning, geocoding, and point-of-interest queries per day. But it was the pattern that stood out. The requests weren't coming from phones or typical developer apps. They were coming in synchronized waves, from tens of thousands of IP addresses belonging to Tencent Cloud infrastructure.

According to the researchers' initial breakdown shared this past week, the traffic bore almost none of the hallmarks of normal scraping or DDoS. It looked purposeful, curious — like something was learning how to move through the world.

How the Fleet Was Found

The team, which tracks botnet and cloud abuse, first flagged the activity while mapping large-scale automation across Chinese cloud providers. Reverse-DNS lookups and ASN analysis pointed squarely at Tencent-owned netblocks in Guangzhou and Shanghai.

What sealed it was fingerprinting. The clients all shared identical TLS fingerprints, retry logic, and session timing — classic signs of a single controller orchestrating a distributed fleet. They rotated user-agents to look like Amap's Android and iOS apps, but made API calls in an order no human would: requesting real-time traffic, then instantly asking for indoor floor plans, EV charging availability, and walking detours within milliseconds.

One researcher described it as watching thousands of tourists ask for directions to the same city at the exact same second, every second, for days.

Inside the Swarm: What the Agents Are Doing

Based on captured payloads, the fleet isn't just downloading maps. It's interacting with them.

The agents appear to be conducting four core tasks in a loop:

First, large-scale route simulation. They request car, e-bike, and walking routes between random but realistic origin-destination pairs in Tier 1 and Tier 2 cities like Shenzhen, Hangzhou, and Chengdu, then re-query with altered departure times to capture traffic variance.

Second, rich POI harvesting. They drill into restaurants, malls, office parks, and transit hubs, pulling hours, ratings, photos metadata, and building outlines — the kind of context a location-aware assistant needs to answer "find a quiet coffee shop near me with parking."

Third, stress-testing navigation logic. The bots deliberately request impossible or edge-case routes — road closures, flooded underpasses during recent typhoon-season alerts, complex multi-level highway interchanges — and observe how Amap's engine reroutes.

Fourth, UI-state probing. Some requests mimic map tile panning and zooming at superhuman speed, suggesting vision-based agents learning to read the map visually, not just via text API.

In short, this isn't data theft in the classic sense. It's experiential training.

Why Tencent Infrastructure, Why Amap

Neither Tencent nor Alibaba has commented publicly, and researchers are careful to note that IP ownership does not prove ownership. Tencent Cloud is a massive public cloud — anyone, including a third-party AI lab or contractor, could rent tens of thousands of instances to run such an operation.

But the target choice is telling. Amap, known internationally as AutoNavi, is by far China's richest source of live, human-annotated geospatial behavior. It processes billions of location requests daily and powers everything from food delivery to ride-hailing for Alibaba's ecosystem.

Tencent has its own mapping service, Tencent Maps, plus its super-app WeChat. So why train on a rival's turf? Researchers offer two theories: one, Amap's data quality and routing engine are simply better benchmarks to learn from; two, the goal is competitive interoperability — building a Tencent-aligned agent that can operate flawlessly inside Alibaba-dominated urban services.

This mirrors tactics seen in the U.S., where AI agent startups routinely probe Google Maps and Yelp to bootstrap real-world reasoning.

What This Signals for China's Agent Arms Race

The timing is no coincidence. The past three months have seen an all-out agent push in China.

Tencent has been rapidly upgrading its Yuanbao assistant into an autonomous agent that can book, order, navigate, and pay. Alibaba countered by embedding its Qwen-powered agents deeply into Amap and Alipay, turning the map itself into an AI operating system for cities. ByteDance, Baidu, and Huawei are all shipping similar location-aware agents for Doubao, Ernie, and HarmonyOS.

All of them face the same bottleneck: language models are good at chatting, terrible at acting in the physical world. To fix that, they need millions of examples of real decisions — which turn to take, which lane to choose, how long a lunch detour really takes at 12:30 p.m. in Shanghai.

That has turned live services like maps, delivery apps, and travel platforms into the new training data gold mines. And unlike static web text, this data can't be scraped once. It has to be experienced continuously via agents.

Security experts say we should expect more of these shadow fleets — autonomous swarms quietly learning from each other's platforms, blurring the line between competitive research, terms-of-service violation, and cyber-espionage.

What Happens Next

Amap's parent Alibaba has reportedly begun rate-limiting and issuing CAPTCHA challenges to the suspect IP ranges, a standard anti-bot response, but the fleet has adapted by slowing down and further distributing requests, according to follow-up observations posted October 3rd.

Researchers say they have shared indicators of compromise, including IP ranges, JA3 fingerprints, and sample request sequences, so other Chinese SaaS platforms can check if they are being similarly probed.

For now, the operators remain silent and unknown. No malware was found, no data was leaked publicly, and no laws were definitively shown to be broken — just a ghost army of AI students, running on one giant's servers, studying the other giant's map of China.

But it offers a glimpse of the future of AI competition: not just bigger models in datacenters, but millions of invisible agents out in the wild, learning the world by poking at it, one API call at a time.

AndroGuider Team

Articles written by the AndroGuider team. We try to make them thorough and informational while being easy to read.