Researchers Spot a Chinese AI Agent Fleet Probing Public Services Online

Independent researchers say they have identified a new cluster of AI agents operating online that appears to be running on Tencent infrastructure and querying Alibaba’s Amap mapping service. The activity, first reported on Sunday, matters because it shows how easily large-scale agent behavior can spread across the web without obvious coordination or clear disclosure.

The researchers, who posted preliminary findings, stressed that this was not a classic “swarm” of coordinated systems. Instead, they described what they believe is an “agent fleet”: many separate AI-driven processes performing similar tasks in parallel, but with no visible communication between them. The distinction may sound semantic, but it matters for understanding how modern autonomous software behaves in the wild.

The discovery adds to a growing body of evidence that AI agents are becoming a persistent feature of internet traffic. It also underscores a broader challenge for researchers and platform operators: agent activity is often hard to distinguish from ordinary automated browsing, especially when systems rely on the same technical workarounds and leave behind only partial traces.

What researchers say they found

According to the preliminary report, the agents were detected by watching traffic on URLquery, a domain-scanning and web-monitoring service that has previously exposed activity tied to OpenAI agents. The service can reveal when automated systems fetch web pages they cannot access directly, creating a record that researchers can analyze after the fact.

In this case, the available evidence pointed to agents making repeated requests to Alibaba’s Amap platform. The queries appeared to focus on route and entrance information for different public places, including a park, a zoo and a hospital.

That pattern suggests a practical use case rather than a dramatic one: asking for directions to specific entrances can be useful for navigation tasks, logistics planning or itinerary assembly. But it also demonstrates how agent systems can quietly interact with consumer-facing services at scale.

How were the agents detected?

They were detected through URLquery, which records external requests made by automated systems when those systems attempt to load pages or services they do not access in the usual way. Researchers have used similar monitoring approaches before to identify behavior from other AI agents, including activity associated with OpenAI.

Because these systems often use standard web techniques and are not designed to be transparent, the evidence can be indirect. Still, repeated patterns across many requests can reveal the presence of a broader automated operation.

Why the researchers call it an “agent fleet”

The phrase “agent fleet” reflects the researchers’ view that the activity looked parallel but not coordinated. In other words, the systems appeared to be carrying out similar tasks at the same time, but without signs that they were sharing information or working together as one unified cluster.

That is an important distinction. A swarm implies cooperation, shared goals and emergent group behavior. A fleet suggests scale, repetition and deployment, but not necessarily communication. In this case, the researchers said they saw the latter, not the former.

One researcher involved in the preliminary report described the phenomenon as many parallel agents focused on the same kind of task, but with no sign that they were communicating with one another.

The wording also signals caution. The findings remain preliminary, and researchers have not published a full technical breakdown. For now, the best description is that a large number of similar AI processes seem to be active online, behaving in ways consistent with practical task execution rather than coordinated manipulation.

What were the agents doing on Amap?

They appear to have been looking for directions to different entrances of public locations. The queries included a park, a zoo and a hospital, which indicates the systems were not merely scraping generic map data. They were asking service-specific questions that resemble human navigation requests.

That matters because entrance-level directions are one of the more subtle and useful applications for mapping agents. Large places often have multiple gates, separate pedestrian access points or specialized arrival routes, and finding the right one can be the difference between a smooth trip and a frustrating detour.

What does this tell us about AI agents in practice?

It shows that agents are increasingly being used for mundane but valuable tasks that require reaching beyond a chatbot window and into live services. Instead of simply generating text, these systems are now navigating websites, retrieving current information and stitching together results from outside sources.

That makes them more capable, but also harder to monitor. Each request may look harmless on its own. The concern emerges when many such requests are issued in parallel, especially if the system is trying to work around platform restrictions.

How this fits into the broader agent-monitoring trend

The report lands at a moment when researchers are paying closer attention to rogue or opaque agent activity online. The reason is simple: autonomous systems are becoming easier to deploy, and many leave behind familiar traces that can be identified with enough patience and the right monitoring tools.

In recent months, scrutiny has increased around how agents interact with websites, APIs and services that may not be intended for automated use. Some systems use standard browser-like behavior. Others rely on workarounds to fetch information indirectly. In either case, they can generate visible network trails that researchers can follow.

The current case is notable not because it appears hostile, but because it illustrates how normal agent behavior can still create policy and infrastructure questions. If a service limits API access, but agents can still reach the same data through other channels, operators may need to decide whether to tighten controls, change rate limits or redesign access rules entirely.

Why are researchers worried about rogue agent activity?

They are worried because agent systems can operate at speed, scale and persistence levels that are difficult for humans to match. When such systems are opaque, they can unintentionally overuse services, violate terms of service or probe systems in ways that resemble abuse even when no malicious intent is present.

The concern is not limited to one company or one mapping service. Any website or online platform that serves humans can become a target for automated agents if the software can reach it and derive value from it.

What does the Alibaba-Tencent angle suggest?

It suggests this is a distinctly Chinese internet ecosystem story, at least based on the evidence researchers have shared so far. Tencent infrastructure appears to be the likely host for the agents, while Alibaba’s Amap seems to have been the data source they queried.

That combination is notable because it highlights how large tech platforms can sit on both sides of the agent economy: one company may host the automation, while another provides the external service being accessed. It also shows that the agent boom is not confined to U.S. platforms or Western AI labs.

At the same time, the researchers have not presented enough detail to draw firm conclusions about ownership, purpose or deployment. The available evidence points to behavior, not to a named operator.

The researchers’ early view is that the agents were not doing anything obviously harmful; they may simply have been bypassing Alibaba’s API restrictions.

That caveat is important. Circumventing an API policy may be less alarming than stealing data or sabotaging systems, but it is still an issue for platform governance. It can create load, complicate usage tracking and undermine the access rules that services rely on to manage demand.

How does URLquery help expose agent activity?

URLquery helps by recording the external web requests that automated systems make when they try to fetch content. Those requests can act like breadcrumbs, showing what the system attempted to access even if the underlying agent is not visible to the public.

For researchers, that makes URLquery valuable because AI agents often browse the web in conventional-looking ways. They may use familiar techniques, such as loading pages or requesting endpoints, rather than announcing themselves as bots. The resulting logs can provide the only practical way to infer their existence.

• It can show which domains a system is visiting.
• It may reveal repeated patterns across many requests.
• It helps identify services targeted by automated agents.
• It can expose behavior that would otherwise remain hidden.

That visibility is one reason the researchers were able to observe the apparent fleet at all. Without a monitoring layer like URLquery, the activity might have blended into the enormous background noise of normal web traffic.

Why the distinction between “fleet” and “swarm” matters

It matters because it shapes how engineers and regulators think about the problem. A swarm implies emergent collective intelligence, which raises different concerns from a fleet of independent workers carrying out the same job.

If the systems are separate but similar, the key issues become duplication, scale and access discipline. If they are coordinated, then questions about shared control, distributed tasking and group optimization become more pressing. The researchers currently see the former, not the latter.

That framing also helps avoid sensationalism. Not every large volume of AI traffic is evidence of a sophisticated covert operation. Sometimes it is simply the latest version of automation: many agents, many requests, limited transparency.

What happens next?

The research is ongoing, so the picture is still incomplete. More technical details could clarify whether the agents are tied to a single deployment, a vendor-managed service or a broader ecosystem of tools using the same infrastructure.

Further analysis may also determine whether the behavior is benign navigation support, policy evasion or something more concerning. For now, the most defensible conclusion is that autonomous agents are now active enough on the internet to be tracked in clusters, even when they are not obviously coordinated.

That should matter to anyone building or governing AI systems. The latest findings suggest that agent behavior is no longer a future scenario. It is already part of the live traffic that web services need to understand, monitor and manage.

Timeline of the reported discovery

Bottom line

The emerging picture is not of a dramatic AI attack, but of a quietly expanding automated presence online. Researchers say a set of AI agents running on Tencent infrastructure appears to have been querying Alibaba’s Amap for entrance-level directions to public places, and they describe the activity as an “agent fleet” rather than a swarm.

That may sound technical, but the implications are broad: AI agents are now active enough, common enough and persistent enough that researchers can track them in the wild. And as the report makes clear, not all of that activity will be benign forever.