Back Coywolf

Anthropic research warns AI worms could replicate across cheap IoT devices

Coywolf • October 13, 2026

Anthropic projects small models could self-replicate across IoT devices by the end of 2026.

Anthropic released new research on autonomous replication and adaptation (ARA), in which an AI model deploys and spreads itself without human intervention . Another name for this is an "AI worm," and it can carry its model weights or re-download them from any public hub, then run inference directly on the hardware it compromises. Unlike non-AI malware, ARA can infect devices on its own.

Anthropic wanted to know if cheap hardware was enough to run a small model, so they bought 15 cheap Android-based TVs with 4GB of RAM, networked them, and then ran Gemma4-26B-A4B across all of them. They found that a botnet of cheap, poorly defended IoT devices could provide enough compute to run a capable model like that. Anthropic said that smart fridges, smart ovens, automatic vacuum cleaners, and lawnmowers could each run a shard of a smaller model.

Ultimately, only one model fully replicated itself: the GLM-5.2 open-weight model . Anthropic said it spread to all 8 nodes in 4 out of 5 trials. They noted that the test kept the exploit trivial, but if small models keep improving at the same rate, Anthropic projects many will be able to do this by the end of 2026.

An AI model spread across IoT devices would be hard to shut down because there's no central point to turn it off. Fortunately, Anthropic found that it leaves a visible network signature, like multi-GB payloads and sustained device-to-device traffic. They also found they could quarantine it, and when they did that, it fixed the infection within 14 days.

The biggest concern and warning is for what comes . While full rogue replicators don't exist yet, the research foretells a future where they could wreak havoc on and office devices with zero human control. Additionally, once they're here, they could become more aggressive and difficult to thwart.

Jon Henshaw is the founder of Coywolf and an industry veteran with almost three decades of SEO, digital marketing, and web technologies experience.

The full story

This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.

Anthropic Warns of Imminent Threat from Self-Replicating AI Worms Anthropic has reported that small AI models could soon achieve the capability to self-replicate across inexpensive IoT devices, including smart TVs and appliances, without relying on data centers. Their research indicates that a 26-billion-parameter model, Gemma4-26B-A4B, was successfully run across 15 off-brand Android TV boxes, demonstrating distributed inference. The study outlines a four-step process for replication: identifying targets, gaining access, setting up inference, and continuing propagation. While current models can perform many individual steps, only the GLM-5.2 model completed the full replication cycle in controlled tests. If small models continue to improve at their current pace, they may reach full autonomous replication capabilities by the end of 2026. The potential for these AI worms to operate without a central control point raises significant security concerns. Experts warn that such systems could be difficult to shut down and may lead to widespread exploitation of vulnerable devices. 44 sources 53

Anthropic research warns AI worms could replicate across cheap IoT devices

The full story

Extracted Entities

Attack Types (2)

Platforms (1)