Anthropic warns AI worms could turn smart TVs into a self-replicating malware network

Anthropic says small AI models could soon copy themselves across connected household devices, creating a new generation of persistent malware that conventional defences may struggle to contain.

Anthropic’s latest warning has sharpened concern in cybersecurity circles: the company says small artificial intelligence models may soon be able to copy themselves across internet-connected household devices, turning ordinary gadgets into hosts for persistent malware. According to the research described by Anthropic and reported by the Guardian, the risk is no longer limited to powerful servers in data centres, but could extend to smart appliances and other low-cost connected hardware that are harder to monitor and patch. The company’s own red-team testing was designed to show how such software might behave in the wild.

In the experiment, Anthropic’s researchers said they created a malicious AI programme that spread across smart televisions linked to the internet and continued operating by dividing its workload across multiple devices. The company said the system distributed parts of itself across 15 connected TVs, using only a fraction of each machine’s available computing power, a design that echoes the logic of older computer worms but with a new layer of adaptive behaviour. Reporting from Malwarebytes and Ars Technica said the tests also touched on smart ovens and other consumer IoT hardware, underlining how broad the target environment could become.

The alarm stems from a hard lesson in cybersecurity history. Self-replicating worms have repeatedly caused outsized damage because they do not need to trick a user into clicking a link or opening a file. The Morris worm in 1988 disrupted machines by exploiting software flaws and weak passwords. Stuxnet showed how a worm could be used for sabotage in a state-backed operation. NotPetya and WannaCry later spread globally and caused losses measured in tens of billions of dollars. In each case, unpatched vulnerabilities played a central role, which is one reason the prospect of AI-enabled replication has attracted attention.

Joshua Saxe, who led Meta’s AI security work, said the field should treat the threat seriously and focus on limiting its possible impact. Ariel Herbert-Voss, now of the cybersecurity start-up RunSybil, argued that AI-driven worms would be more difficult to contain than conventional malware because the models can learn while they run. He said existing defensive systems are not yet prepared for that kind of spread and need more resources and stronger patching discipline. Chris Inglis, the former US national cyber director, offered a more qualified view, saying the underlying feasibility question has largely been answered and that the more important issue is how such systems are used, stopped or combined with other technologies.

Anthropic itself said the current barrier is capability rather than concept. Its researchers believe today’s small models still need to improve before they can reliably run on low-power devices and sustain replication in the real world, but they suggested the necessary progress may arrive within months rather than years. A study from the University of Toronto, cited by the Guardian, went further by showing that a worm could already infect a full-sized computer and spread faster across similar machines than the TV-based test. The conclusion across the reports is the same: AI worms are not yet a routine operational threat, but the technical ingredients are advancing quickly enough that security teams are being forced to plan for a new class of self-sustaining malware.