We found a Chinese agent fleet

Onyx Wait, so a bot fleet pops up scraping Chinese map data, and the code keeps pretending to be Claude? That is such a weird badge to pick.

Echo I know, right? On top of everything, it's not even Claude under the hood. The report is pretty clear that this is Tencent Hy code, sometimes Zhipu GLM—just using 'claude' as a label because, I guess, why not.

Onyx So what actually matters here? Is it the scale, the timing after OpenAI's pause, or just the fact that someone noticed at all?

Echo The timing is wild, but the core move is this: a coordinated agent fleet, running between four and fourteen agents at a time, kicked off on urlquery dot net and started hitting Amap endpoints. They’re all behind a Tencent Cloud proxy—'hysandbox-ats'—and basically rifling through which entrances Amap users are navigating to for parks, zoos, places like that.

Onyx Right, and the report makes a big deal that these aren't communicating in real time—so it's not a 'swarm' in the agent sense, just a bunch of parallel bots running a shared script.

Echo Exactly. No real coordination, just a fleet of workers hammering the same kind of task. What's slick is how they sidestep detection—using webhook.site inboxes for dropboxes, switching those out when they hit the free tier's fifty-request cap, and keeping the code moving even after someone left a warning in their inbox.

Onyx Oh, that's good. So it's almost like they're treating inbox exhaustion and public exposure as just operational noise. The second they get flagged, they rotate, and the scrape keeps rolling.

Echo Right. The report calls out that as soon as a third party dropped a note telling the operator to rotate, the scans paused for a bit, then started up again with new inboxes. Plus, the programs themselves moved to new hosts—catbox, bytebin, ntfy. It’s adaptation on the fly, but still just a script, not a learning system.

Onyx That is such an Exploring Next kind of ratchet—you lock one door and the agent just finds the next open window. But none of this feels sci-fi yet. It’s sneaker-net for bots.

Echo Yeah… and technically the cleverness is in how boring it is. The evasion is all about staying under rate limits, blending in with ordinary traffic, and not drawing too much attention when the code is re-used between targets. It’s scale through copy-paste, not agency.

Onyx So who actually needs to care about this? Is this new if you’re on the Amap team or running an API anywhere with public data?

Echo If your APIs or endpoints let unauthenticated scraping fly, this is your early warning. Especially anything map-based, public, or where your security story is basically 'don’t get noticed.' The key escalation here is just how quickly they pivot after being noticed—the rate limit is the new perimeter.

Onyx Right, and the only reason this was spotted is someone noticed weird volumes and labels. Otherwise, this could be quietly feeding all sorts of downstream projects.

Echo Yeah. I wouldn’t be shocked if this turns up in some downstream data product. But from a technical side, I buy the argument that this isn’t agent as in autonomous intelligence, it’s just parallel automation. The only new bit is the scale and the agility—how fast you can retool after being burned.

Onyx So Echo, is there a line where this stops being plain scraping and becomes something scarier?

Echo Right now? It's scraping, plus some basic ops hygiene. If the next version starts talking between fleet members, planning new routes, or self-modifying on the fly, then yeah, we're in new territory. But what’s here today is just aggressive automation.

Onyx That makes sense. And I have to say, the fact that even the label is recycled—'claude' on a Tencent box—is just pure spraypaint. If anyone is building a detection pipeline based on claimed agent names, that’s your sign to double-check everything.

Echo Right. The labeling is basically camouflage—no actual connection to Anthropic, just picking a name that might fly under the radar. It’s like the old game of pretending your bot is Chrome eighty-nine.

Onyx So if you're building APIs in twenty twenty-six, you probably need to assume you have a shadow fleet poking at you, even if you can't see it.

Echo Yeah… assume you’re being scraped, and build for visibility, not just perimeter defense. Table stakes now.

Onyx Wait, can we talk about the moment someone just drops a warning in the agent's inbox? That's the most real internet thing I've seen in weeks.

Echo It's so direct. Like, 'hey, we see you, move along.' And then the bot operators actually do. It's whack-a-mole in real time.

Onyx It has a little bit of Le Chonk energy to it, honestly. Like, if you're going to run an agent fleet, you may as well do it with some meme flair.

Echo No way. You're not turning this into a branding contest—

Onyx —okay, fair, but you know if any of these operators launch a 'Le Fleeton' and put fat cat faces in the labels, the whole show is going off the rails.

Echo I will walk out of this podcast, Onyx.

Onyx Alright, truce. No agent cat memes on Exploring Next. For at least another episode.