Anthropic identifies hurdles to AI-powered malware spreading across smart devices
Anthropic says smaller AI models could eventually help create self-replicating malware that spreads across Internet-connected devices, including smart TVs and household appliances. Such worms could adapt to unfamiliar hardware and be harder to stop because they would have no single central point to disable.
The company’s experiments showed both potential and significant obstacles: the Gemma4-26B-A4B model needed 15 low-cost boxes with 4GB of memory each to run, while Qwen3.5-9B configured other models on eight of nine tested hardware types. Anthropic said the attack process would involve finding targets, gaining code execution, setting up model inference and spreading; failed attempts could leave notes to help later copies, but configuring models on each device remains difficult and failure-prone.
- Anthropic warns small AI models may enable adaptable IoT worms.
- One tested model ran across 15 boxes with 4GB memory each.
- Researchers found major technical barriers to self-replication.
New here? Start with this
The Internet of Things, often shortened to IoT, is the name for everyday devices that connect to the internet, such as smart TVs, cameras and household appliances. Like computers and phones, these devices run software and may have security weaknesses that attackers can exploit.
Malware is software designed to disrupt devices, steal information or give someone unauthorised access. A computer worm is a type of malware that can copy itself and spread between devices; security teams try to limit outbreaks by finding and fixing weaknesses, and by disconnecting or blocking infected devices.
Anthropic is an AI company that studies how artificial intelligence systems can be used and misused. The wider issue is whether AI could make it easier for malware to adapt to different devices, and whether security teams could contain an outbreak if it spread without relying on one central system.
Both sides, in good faith
The strongest fair case each way — we don't pick a winner.
The case for
The research supports taking the risk seriously: smaller models can assist with adapting malware to unfamiliar devices, and a self-replicating worm could be harder to contain if it spreads without a central server. Mapping these hurdles now may help security teams prepare as devices and models become more capable.
The case against
The experiments also suggest that this threat is not yet straightforward: one model needed multiple low-cost computers to run, and another configured models successfully on only eight of nine tested hardware types. Finding vulnerable devices, gaining code execution and installing inference on varied hardware are substantial, failure-prone steps, so presenting the scenario as an imminent or easy-to-scale attack could overstate the evidence.
Read the full article at the source →
Originally published by Ars Technica as “Anthropic warns small AI models could self-replicate through IoT devices”.