Full Breakdown
Anthropic Warns AI-Powered Worms Could Turn Smart Homes into Botnets
By Drooid · · How we work
Core Event: Anthropic’s AI-Worm Research Highlights Emerging IoT Threat
Anthropic’s frontier-red-team released a report indicating that small AI models may be “months away” from self-replicating across internet-connected devices such as smart ovens and off-brand Android TV boxes. In lab tests, the researchers distributed a model across 15 TV boxes, infiltrated a simulated device, and in four out of five trials spread to all eight target devices. The company warns that such “AI worms” could form decentralized botnets that lack a single control point.
Background & Context: Growing Smart-Home Footprint and Recent Cyber Incidents
Industry association Bitkom reports that roughly 52 % of German households now use smart-home applications. The vulnerability of such devices was underscored by a recent cyberattack on French supermarkets that compromised refrigeration control systems, causing temperatures to climb to 50 °C and spoiling food. France’s cybersecurity agency CERT-FR confirmed the intrusion but found no evidence of an AI worm.
Data & Statistics
- 15 off-brand Android TV boxes (? $55 each) ran the Gemma4-26B-A4B model at 0.39 tokens / s.
- In a separate test with the GLM-5.2 model, replication succeeded on all eight nodes in four of five trials.
- A 9-billion-parameter model ran on 8 of 9 hardware types tested.
- Anthropic estimates a model compact enough for household hardware could reach GLM-5.2-level capability by the end of 2026 if historical improvement rates continue.
Official Statements & Responses
Anthropic emphasized that the experiments deliberately removed major digital safeguards to illustrate a ceiling rather than a forecast. CERT-FR advised companies to regularly update devices, install security patches and, where possible, disable unnecessary internet access.
Verbatim Quotes
- “Our experiments suggest that small AI models may be months away from being capable enough to self-replicate on commercial internet-of-things (IoT) devices,” — Anthropic
- “ "A growing number of home and business devices, such as security cameras and video-conferencing units now ship with small built-in AI chips, designed to run AI models on the device itself," Anthropic wrote.” — Anthropic
- “What’s new here is that these models were much smaller, in terms of parameter count, than previous ones which have been shown to be capable of self-replication,” — Jeffrey Ladish, Palisade
Conflicting Reports & Gaps
- No concrete evidence links the French supermarket attacks to AI-driven malware.
- Anthropic used an unauthenticated Android Debug Bridge port, making the initial compromise “deliberately easy,” so real-world feasibility remains unproven.
- The projected 2026 timeline is based on a “general capability index” and is described as a “rough signal only,” leaving uncertainty about when such worms might appear.
What’s Next: Anticipated Developments
Anthropic projects that models with 30 billion parameters or fewer could achieve full autonomous replication by the end of 2026 if improvement trends persist. The company suggests defenders monitor for unusual network activity—large model downloads, unexpected inter-device communications and scanning behavior—to detect early infections.