Anthropic is issuing a warning about future AI-driven malware targeting IoT devices like smart fridges and vacuum cleaners. They predict that by next year, small AI models could autonomously self-replicate across various internet-connected devices. This capability could lead to advanced computer worms that spread without human intervention and adapt to unfamiliar hardware. Anthropic's recent experiment showed that while creating such malware is currently challenging, it is not impossible. The potential for AI malware is concerning because it would be extremely difficult to shut down due to its decentralized nature. Unlike traditional malware, these AI worms could independently discover and exploit new devices. Researchers found that small AI models can run on low-cost IoT hardware by using techniques like quantization and workload sharding. This allows the models to load their weights and generate outputs on these limited devices. The next research step explored if these small models could create self-replicating infections similar to existing IoT worms. However, AI worms could be far more dangerous by dynamically customizing exploits and payloads for specific targets. Once initiated, these AI worms could prove harder for security teams to eradicate. The AI model weights themselves could become the malware payload, accompanying the replication process.