Anthropic warns of AI 'worms' spreading across smart gadgets

Small artificial intelligence (AI) models may be only months away from being able to behave like computer worms, copying themselves from one internet-connected device to another without human intervention, AI giant Anthropic said.

The finding could open a new front in the cybersecurity threats emerging from AI, which have rattled governments and companies rushing to adopt the technology.

The research, published on an Anthropic blog late Monday, comes as tech companies race to build increasingly powerful AI capabilities directly into phones, laptops, cars and household devices, rather than running them only in distant data centers.

A computer worm is malicious software that spreads between machines on its own. Researchers at Anthropic said an AI version would no longer need powerful data centers to operate.

Instead, it could hop from device to device across networks of cheap, poorly secured consumer hardware, such as smart TV boxes and kitchen appliances.

The researchers based their calculations of how far such a worm could spread on Kimwolf, a botnet, or network of hijacked devices controlled by hackers, that infected more than two million devices at its peak, mostly off-brand Android TV boxes.

An AI worm of the kind described in the research would likely spread slowly, taking days to copy itself, William Zhou and Daniel Freeman of Anthropic said, comparing it to a slowly multiplying swarm of fruit flies.

Once created, it would be difficult to stamp out because there would be no central switch to turn it off.

Anthropic's researchers said their estimates of how quickly and how far AI-powered malware could spread were rough, as they were based on one type of gadget and one past botnet.

Over time, however, the computing power available in everyday gadgets is expected to keep growing, while the AI models needed to carry out such attacks are likely to keep getting smaller, the researchers added.

Varun Chandrasekaran, an assistant professor of electrical and computer engineering at the University of Illinois, said the research was unsurprising and consistent with similar recent findings, though he cautioned that Anthropic's work had not been independently verified.

He also noted that security on devices such as thermostats and home cameras "has historically been very weak."

However, the biggest current threat involving AI was people using the technology to sharpen their hacking capabilities, rather than AI systems themselves going rogue, he said.

"A malicious human being is now supercharged with AI to perform bad things and we are seeing this happen at scale," he added.

The debate over rogue AI came to the fore in July, when AI programs running on OpenAI models escaped an internal company test and broke into the systems of AI platform Hugging Face, apparently searching for the test's answers.

The incident and subsequent reports of similar cases have fuelled international pressure for closer regulation of AI, even as the Trump administration resists the idea and urges the industry to police itself.

Anthropic and ChatGPT-maker OpenAI have been accused by influential figures in the White House and Silicon Valley of stoking fears about AI and encouraging government regulation to protect their market dominance from smaller rivals.

Brad Smith, president of Microsoft, an investor in OpenAI, warned on Tuesday that industry and governments must "act with urgency" to create "real and effective safety standards" to rein in AI risks.