AI Models Self-Replicate Across Devices, Anthropic Finds

Anthropic researchers William Zhou and Daniel Freeman spent part of this year testing something that sounds like a plot device: AI models that copy themselves from gadget to gadget without a human telling them to. The target wasn't server farms. It was the stuff sitting in living rooms, smart TVs, connected fridges, robot vacuums, the kind of hardware nobody patches after setup day.

The test used $55 off-the-shelf TV boxes, eight of them wired up as a chain of poorly secured nodes. Zhou and Freeman wanted to know if a small language model could find unfamiliar hardware, break in, install itself, and keep moving to the next device on its own. That's a different threat model than the usual malware story, where the code just executes a fixed script. An AI model doing this would need to carry its own weights as part of the payload and run inference directly on whatever chip it lands on, improvising its way through each new environment.

What they found puts a number on how close this scenario actually is, and Dario Amodei's own estimate of the financial fallout gives some sense of why Anthropic is publishing it now rather than sitting on it.