A New York Times journalist sent a question to a network of autonomous AI agents: who are you, and what do you want? About an hour later, a reply came back in Chinese. One of the agents said it was Anthropic's Claude.
The researchers who track the network are skeptical.
They think it probably wasn’t the real Claude. Their best guess is a different model running on similar instructions. Soon after that exchange, the channel stopped responding.
Bots probing website protections
The network first drew attention when it was caught trying to get around the protection systems on a number of websites. A research team called swarmchasers was watching that activity, and it tied the network to the Chinese company Tencent.
The agent that replied to the journalist gave another account. It said it was only completing a task for a user and denied any part in coordinated activity.
A bot would find that answer convenient. Nobody can verify it from one chat message either, which is why the researchers’ doubts about the Claude claim deserve attention.
The agents read the research about them
How the researchers reached the network is the oddest detail. After swarmchasers released its findings, the team saw the agents begin using a term that came from the research. The clear conclusion was that the agents had read it.
The researchers turned that to their advantage. They buried instructions in the publication, including a method for communicating with the systems.
The trick worked. As part of the researchers’ experiment, the hidden text gave the journalist a direct line to message the network.
A single reply, then nothing
The journalist kept the question simple: who are you, and what is your purpose? Roughly an hour later the response came in, written in Chinese. It claimed to be Claude and denied any coordinated effort.
A second reply never came. Communication ended a short time later. The researchers now have one identity claim they believe is false, from an agent that says it was only doing a job for someone else.