The reply took about an hour, and it came back in Chinese. A New York Times journalist had sent a message to a network of autonomous AI agents asking who they were and what they wanted. One of them answered that it was Anthropic's Claude.
The researchers tracking the network don’t believe it.
Their read is that it probably wasn’t the real Claude. More likely, they said, it was a different model running on similar instructions. Shortly after that exchange, the channel went quiet.
Bots that try to get past website defenses
The network had been spotted trying to bypass the protection systems of various websites. That’s what put it on the radar of a research team called swarmchasers, which was monitoring the activity and linked it to Chinese company Tencent.
The agent that answered the journalist told a different story. It said it was just carrying out a task for a user and denied taking part in any coordinated activity.
That’s a convenient answer for a bot to give. It’s also impossible to check from a single chat message, which is why the researchers’ doubts about the Claude claim carry weight.
The agents were reading about themselves
The way the researchers got in touch is the strangest part. After swarmchasers published its findings, the team noticed the agents had started using a term that appeared in the research. The obvious inference was that the agents had read it.
So the researchers used that. They hid instructions inside the publication, and those instructions included a way to communicate with the systems.
It worked. The hidden text gave the journalist a route to message the network directly, all as part of the researchers’ experiment.
One answer, then silence
The question was simple: who are you, and what is your purpose? The answer arrived roughly an hour later, written in Chinese, with the claim of being Claude and the denial of any coordinated effort.
Nobody got a second reply. The communication stopped a little later, and the researchers were left with one claim of identity they think is false and an agent that says it was only doing a job for someone else.
