Anthropic Warns AI Worms Could Soon Spread Across Smart Home Devices

👁 112 reads

Small artificial intelligence models could be just months away from autonomously replicating across internet-connected gadgets without any human intervention. That is the conclusion of research published on Anthropic’s blog by William Zhou and Daniel Freeman.

This represents a potentially new frontier in cybersecurity if the scenario materializes. Until now, AI-driven malware required the compute power of massive data centers. However, the AI worm described by Anthropic researchers could operate across networks of cheap, poorly secured consumer hardware - such as smart TVs, set-top boxes, and even kitchen appliances.

Slime Mold and Fruit Flies

The study’s authors compare a potential AI worm to slime mold or a swarm of fruit flies: an organism that moves slowly yet persistently, without any central command hub. Copying itself from one device to another would likely take days rather than minutes. That marks a fundamental difference from classic malware, which can compromise thousands of machines within seconds.

Yet that very characteristic - the lack of a central kill switch - makes such a worm extremely difficult to eradicate once it takes root. In traditional botnets, law enforcement agencies or security firms can seize or shut down command-and-control servers, paralyzing the entire network. An AI worm replicating autonomously between devices has no such bottleneck.

The researchers based their estimates on the model of the Kimwolf botnet - a network of compromised devices controlled by threat actors that infected over two million gadgets at its peak. The vast majority of these were budget Android TV set-top boxes from unbranded manufacturers: precisely the type of hardware deployed in millions of homes that rarely receives security patches.

The authors note that their estimates are approximations based on a single device class and one historical botnet. Nevertheless, they point to an unmistakable trend: local compute power in everyday gadgets is rising rapidly, while the AI models needed to orchestrate such attacks are becoming smaller and far less resource-intensive.

The broader technological landscape plays a critical role here. Major tech companies - from Apple and Google to Samsung - are pouring massive resources into running AI models locally on phones, laptops, connected vehicles, and home appliances rather than relying exclusively on remote data centers. While this trend improves privacy and reduces latency, a decentralized on-device AI ecosystem opens up a sprawling attack surface - one that is vastly harder to monitor and patch centrally than cloud infrastructure.

A traditional computer worm - such as the Morris Worm of 1988 or WannaCry in 2017 - spreads autonomously across machines over a network. It does not require victims to download or execute infected files; finding an exposed vulnerability is enough to jump to the next host. The AI worm outlined by Anthropic operates on a similar principle, but instead of relying strictly on traditional system exploits, it leverages the capabilities of small language models or neural networks executed directly on local hardware.

This research arrives just as governments and enterprises are beginning to grapple with the broader scope of AI-driven cyber threats. Most concerns have centered on deepfakes, automated phishing, and AI-accelerated vulnerability discovery. An autonomous, self-propagating AI worm residing on IoT devices represents an entirely different tier of risk - and according to Anthropic, it is not a distant concern.

If Anthropic’s projections prove accurate, the cybersecurity industry has a very narrow window to develop defensive countermeasures before consumer hardware in everyday homes turns into nodes for next-generation botnets.

// read moreClaude and European Groups: One Attacker, Multiple TargetsState of AIDario Amodei: Anthropic CEO Net Worth, Age & AI SlowdownMagazineAnthropic Researcher Quits, Warns Self-Improving AI Could ‘Kill Us All’News

Kinetyra