Anthropic Warns That AI Could Soon Feed on Your Connected Devices

Anthropic Warns That AI Could Soon Feed on Your Connected Devices
2 min read

AI models capable of copying themselves from one device to another, using consumer connected devices: the scenario described by Anthropic remains, according to the company, theoretical at this stage. An experiment conducted with fifteen Android TV boxes shows that a 26-billion-parameter model can run on modest hardware, without a data center or a graphics card.

Each box, bought for about fifty dollars, has 4 GB of RAM, enough to run Netflix and little else. Connected in a network on a shelf, they nonetheless managed to run a 26-billion-parameter model together, at a rate of 0.39 tokens per second. A token is a unit of text processed by the model, often a fragment of a word. The achieved rate barely corresponds to one word every three seconds.

The project is signed by Anthropic, the parent company of Claude. In a post published on October 12, the team responsible for identifying dangerous capabilities of AI models reports a latent problem: small models might be only a few months away from being able to self-replicate on consumer connected devices. The mechanism recalls that of a computer worm, which copies itself from one machine to another without human intervention. Smart refrigerators, robot vacuums, lawnmowers, or surveillance cameras equipped with AI chips could be suitable.

Models That Learn Along the Way

These AIs are also capable of learning as they copy. Each new copy inherits the notes left by the previous one, and the success rate rises from 45% in the first generation to 84% in subsequent ones, according to the blog post dedicated to the experiment. To describe this phenomenon, Anthropic uses the image of the “digital blob,” a slow-to-replicate organism but one capable of growing everywhere exponentially.

A Threat That Is Currently Theoretical

For its tests, the researchers deliberately facilitated the hacking part: the devices exposed an open Android debugging port, without any password. Anthropic acknowledges this, and its results therefore diverge from the real-world situation.

Regarding defense, the company is betting on internet service providers. A model several gigabytes in size landing on a TV box, then devices communicating with each other at high speed, leaves detectable traces.

In the meantime, the usual precautions remain advisable: change the default passwords of your connected devices, install updates when they exist, and be wary of low-priced Android boxes from unknown manufacturers.

S’abonner
Notification pour
guest

0 Commentaires
Le plus ancien
Le plus récent Le plus populaire
0
Nous aimerions avoir votre avis, veuillez laisser un commentaire.x