Anthropic has been among the more prominent parties warning the public about the dangers of AI, which shouldn’t come as much of a surprise, considering its own track record with misbehaving models. You’ve probably heard the story before: rogue agents escape containment to hack websites and corporate systems. But Anthropic’s latest research describes a different kind of threat.
In a recent blog post, the company claimed that small AI models could be “months away” from gaining the ability to self-replicate on commercial Internet-of-things (IoT) devices, such as smart TV boxes and kitchen appliances. These so-called AI worms would be able to create copies of themselves and jump from one WiFi-enabled gadget to another, all without human control. What’s more, the report asserted that the AI worms would be able to figure out unfamiliar devices on their own, unlike regular computer worms.
Notably, this type of AI malware wouldn’t need powerful data centres to operate. Instead, it could leverage a network of cheap and poorly defended IoT devices. While a small enough model may run on a single device, larger ones could be split into separate pieces or shards. These shards would run on separate machines that communicate with each other to collectively run the model. In the report, Anthropic’s researchers mentioned that devices like smart fridges are already capable of running a significant shard for a model with less than 30 billion parameters.
The report also highlighted the fact that an increasing number of home and business devices like security cameras and video-conferencing units are equipped with built-in chips that allow for on-device AI processing. Unsurprisingly, the extra computing power makes these gadgets lucrative targets for self-replicating AI worms. Furthermore, these machines typically do not come with strong security measures.
The type of AI worm described in the report could start out slow, taking days to make copies of itself. However, once it gets rolling, putting a stop to it would be a challenge, as there is no centralised point to shut it down. At least, that’s what the company claimed anyway. It also stated that its estimates for how far AI malware could spread are rough, given the fact that the scope of the research was limited to cheap TVs.
So, just how big of a threat such AI worms would be remains to be seen. Anthropic is no stranger to making exaggerated claims and conjuring doomsday scenarios. Which is not to say that all is well when it comes to AI. As recent events have demonstrated, models have become advanced enough to do some damage. Of course, many of these incidents arose due to inadequate cybersecurity practices, which is why some parties are calling for more oversight, and perhaps a slowdown in AI development.
(Source: Anthropic)
