🗣“Small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices, like smart ovens and…Android TV boxes,” no longer needing access to data centers, and proving impossible to stop, “since there’s no centralized point to shut off,” new research by Anthropic indicates.
In lab testing, the company found that one simple model was able to “figure out unfamiliar devices on its own” in 8 of 9 hardware types, and that “if small models improve at their historic rate, they’ll match” the capabilities of more complex models’ potential for self-replication “by the end of this year.”
If that happens, it would be the textbook definition of a “worm” – a piece of self-replicating malware that can spread from one device to another without any human involvement.
“Prior work has already shown that AI agents can act like [a worm]: scanning their network, executing an exploit, and copying model weights to a target machine across a GPU cluster. Datacenter GPUs would be a clear, high-value target for such malware. But expensive accelerators aren’t strictly necessary,” with “cheap and typically poorly defended” household devices presenting a perfect alternative.
Tech companies spent more than a decade touting the IoT future of smart appliances, thermostats, lighting, home security, televisions, pet tech, wearables, medical devices, vehicles, and even ‘smart city’ infrastructure (traffic lights, utilities). Now, Anthropic says this could all be hacked and used by a rouge AI, echoing the warnings of industry whistleblowers.
👍 Boost us | Chat | @geopolitics_prime