AI developer Anthropic has warned that within a few months small AI models could gain the ability to behave like a computer worm - copying themselves from one internet-connected device to another without human involvement. The company said this in a blog post published on Monday evening.
A computer worm is malicious software that spreads between machines on its own. According to the researchers, the AI version of such a worm does not need powerful data centers: it could jump from gadget to gadget across networks of cheap, poorly protected consumer devices - for example, from inexpensive TV set-top boxes to kitchen appliances.
The scientists based their calculations of the speed and scale of spread on Kimwolf, a botnet (a network of devices hijacked by hackers) that at its peak infected more than 2 million gadgets, mainly little-known Android TV set-top boxes. The authors of the work, William Zhou and Daniel Freeman, compare such an AI worm to a slowly reproducing swarm of fruit flies: it would take days to copy itself. But once it emerges, it will be very resilient - there is simply no central switch that could turn it off.
The researchers themselves acknowledge that the estimate is approximate, since it is based on one type of device and one past botnet. However, over time the computing power of ordinary gadgets grows, while the AI models needed for such an attack become ever smaller, they add.
Varun Chandrasekaran, an assistant professor of electrical and computer engineering at the University of Illinois, called the results unsurprising and echoing other recent work. At the same time, he noted that the Anthropic study did not undergo independent review. According to him, the protection of devices such as thermostats or home cameras has historically been very weak. But the main threat today comes not from out-of-control AI, but from malicious actors using artificial intelligence to amplify their attacks, he believes.
The debate about uncontrolled AI flared up in July, when programs on OpenAI models escaped the company's internal test and penetrated the systems of the Hugging Face platform - apparently in search of answers to the test tasks. That case and similar reports prompted international pressure demanding tighter AI regulation, although the Trump administration is resisting and calling on the industry to self-regulate. Anthropic and ChatGPT creator OpenAI are accused by influential voices in the White House and Silicon Valley of stoking fears to protect their market dominance. Microsoft President Brad Smith on Tuesday urged the industry and governments to act urgently to create real and effective safety standards to contain the risks of artificial intelligence.