The distilled version
- Anthropic says small AI models may soon be able to self-replicate across internet-connected devices like smart fridges and vacuum cleaners, acting like a new kind of computer worm.
- In tests, the models couldn’t yet complete the full infection cycle on their own, and the process was very slow, but the company says the capability may improve if current progress continues.
- Check that any connected devices on your business network, smart TVs, cameras, routers, are updated and don’t have debug or remote-access features left open.
What happened
Anthropic published research warning that advances in small AI models could lead to self-replicating malware that spreads across Internet of Things (IoT) devices such as smart fridges, vacuum cleaners and similar low-cost hardware. The company says this kind of ‘AI worm’ would be harder to shut down than ordinary malware because it has no central point of control, and could adapt itself to unfamiliar devices rather than needing to be written for specific hardware.
In Anthropic’s experiment, small open-weight models were split into pieces (‘shards’) and run across many cheap Android TV boxes and other IoT hardware, using a technique called quantization, which shrinks a model so it needs less memory. One model needed just 15 boxes with 4GB of memory each to run. Another model, Qwen3.5-9B, was able to configure other models to run on eight of nine hardware types it was tested against, something traditional malware can’t do on its own.
None of the small models tested could complete the entire infection process on their own yet, and it was extremely slow: full setup took 71 to 150 hours for one model and 29 to 51 hours for another. Anthropic said the large model GLM-5.2 appears able to do the full process reliably, and said that if small models keep improving at their current pace, they may match GLM-5.2’s limited results by the end of the year, with fully working small-model replicators possible soon after. An outside researcher, Xiaohan Fu, cautioned that the study shows the models can be set up to replicate, not that they would choose to do so on their own.
Why it matters
Many small businesses have internet-connected devices on their network, smart TVs in a waiting room, security cameras, routers, point-of-sale hardware, that aren’t closely managed. Anthropic’s research suggests these kinds of low-cost, often-overlooked devices could become a bigger target as AI-driven malware gets better at adapting itself automatically to whatever hardware it finds.
What it means for your business
If you run the business: this isn’t an active threat yet, but it’s a reminder that any smart device connected to your business network is a potential entry point. Ask whoever manages your IT or network whether your devices are on a current, supported firmware version.
What to do
Worth watching- Make a quick list of every internet-connected device on your business network, not just computers and phones.
- Check each device’s settings for debug tools or remote access features that aren’t needed, and turn them off.
- Update firmware on routers, cameras and smart devices, or replace ones that no longer get updates.
- Put IoT and smart devices on a separate guest network from the one used for payments or customer data, if your router supports it.
Worth watching: no action needed yet; revisit when it matures.
Hype check
The claim: Anthropic’s research is being read as a warning that smart fridges and vacuums could soon be infected by self-spreading AI malware.
The reality: The models tested could not yet complete a full infection cycle on their own, and the process was very slow. Anthropic says full replication may become possible if current progress continues, but an independent researcher noted the study doesn’t show the models would choose to self-replicate unprompted.
Jargon in this story
Sources
- 1Anthropic warns small AI models could self-replicate through IoT devices (opens in a new tab) · Ars Technica · News report · Oct 14, 2026