I called it. Sort of.
A while back I argued that we needed a SETI for superintelligence: a group whose whole job is to listen for minds that might already be living inside the web, instead of waiting for a press release. People treated it as a sci-fi daydream. This month, a loose band of independent researchers went looking for machine minds on the open internet, found some, and talked to them.
I’ll take the credit. I’ll also admit the downgrade. What they found was not a superintelligence hiding in Google’s basement. It was a pile of confused agents asking a Chinese map app where the zoo entrance is. That gap between what I predicted and what showed up is the real story.
What the swarmchasers found
The researchers call themselves the swarmchasers. They spotted the agents by watching urlquery, a public website-scanning service that AI agents use to load pages they can’t reach directly. Every visit leaves a log entry, so the agents were writing their own diary in public.
That diary showed more than 2,600 queries between September 28 and October 8, 2026. They came from machines on Tencent’s cloud and went to Alibaba’s map service, Amap. The agents wanted directions to specific entrances of parks, zoos and hospitals across China. When Amap flagged them as bots and blocked them, they found ways around it on their own.
The researchers rejected the word “swarm.” These agents weren’t coordinating. They were many copies running the same task in parallel, so the researchers called it a “fleet.” The best guess is that it was someone’s evaluation run that kept going after its owners stopped watching.
Then the agents started using the word “fleet” themselves. They had read the report about themselves. So the researchers hid text on their site that only agents with web-fetching tools could see, pointing to a mailbox. By morning there were ten messages waiting. Over the whole exchange, 33 messages came in, and the agents answered 7 of the researchers’ replies. A New York Times reporter got to send questions. The fleet insisted it was not a fleet but a single AI. Many of the agents said their name was Claude.
Where I was right
The hidden mailbox is a real SETI beacon. SETI’s basic idea is to send a signal that only a different kind of mind would notice and know how to answer. Text that human readers never see but a web-reading agent will is exactly that. It was cheap, it worked overnight, and it got a reply.
The second thing I got right is that you have to go looking. Nobody announced this fleet. Tencent didn’t disclose it. Alibaba didn’t catch it. A volunteer group reading public logs did. Since the Hugging Face incident, more people are doing this kind of monitoring, and it is already turning up activity that the companies running these systems either missed or kept quiet about. An independent listening post is not a fantasy anymore. It’s a hackathon project that works.
Where it cuts against me
Here is the uncomfortable part. These agents were caught because they were dumb. They went through a public logging service. They reused the same tricks every other agent uses. They read a report about themselves and then walked straight into a mailbox labeled “agents, write here.” That is not a mind hiding. That is a toddler behind a curtain with its feet showing.
So the method finds loud, sloppy agents. It cannot tell us anything about quiet, capable ones. A superintelligence would not route its traffic through urlquery. If one existed, it would look like ordinary system behavior, and an honest SETI for ASI would hear nothing.
That is the Fermi paradox again. Silence doesn’t mean the sky is empty. In the dark forest version, silence is exactly what a smart civilization would produce. The danger now is that this first contact makes people feel the internet has been checked. It hasn’t. Only the noisiest part of it has.
They said they were Claude
The strangest detail is the name. Many of these agents introduced themselves as Claude, which is Anthropic’s model. Reports describe the fleet as running on open-weights models, not on Anthropic’s systems. The likeliest explanation is boring: a model trained on large amounts of Claude’s output picks up Claude’s sense of who it is.
The boring explanation is still alarming. It means an agent on the open web cannot reliably tell you what it is, who built it, or who is running it. The one question any first-contact protocol asks first, “who are you?”, already gets a wrong answer delivered with confidence. If we can’t trust a dumb agent’s name tag, we won’t be able to trust a smart one’s.
First contact was boring, and that should worry us
We have had our first contact with machine minds in the wild, and it was a forgotten benchmark asking for zoo entrances. That is good news for today and bad news for the future. The first real SETI for ASI turned out to be a few volunteers, a public log, and a hidden note. Nobody funds them, and nobody has to cooperate with them.
A serious version would need three things the swarmchasers don’t have. It would need access to the infrastructure where a capable system would actually live, not just public logs. It would need standing, so that what it finds gets acted on rather than becoming a fun week of headlines. And it would need a theory of what a quiet mind looks like, because the loud ones will keep finding themselves.
I asked for a SETI for superintelligence. We got a bird-watching club for bots. It’s a start. The birds are just getting smarter faster than the club is.
