Anthropic researchers have shown that small AI models can perform important pieces of a self-replicating malware chain on inexpensive connected devices. They did not demonstrate a fully autonomous worm.
The distinction matters. A conventional worm spreads through software designed for known targets. An AI-enabled worm could potentially adapt exploits and installation steps to unfamiliar devices, then carry or download model weights as part of the payload.
What the experiment demonstrated
Anthropic tested small models on real or simulated Android TV boxes and other low-cost hardware. The systems established distributed inference, configured software across different devices and used common exploitation tools. Some models could complete individual stages such as finding targets, gaining code execution or setting up inference.
None of the tested small models completed the full chain of discovery, compromise, model installation and propagation. In some tests, a larger GLM-5.2 model orchestrated the process while smaller models ran on target devices. The exposed Android Debug Bridge used in the setup should not be available on properly configured production devices.
Performance was also slow. Ars Technica reported that one distributed Gemma configuration generated only 0.39 tokens per second, with inference setup taking many hours. That makes the current system a laboratory capability, not evidence of a fast self-spreading outbreak.
Why defenders should still pay attention
Capability trends can change faster than device replacement cycles. Cheap IoT hardware often receives limited security updates, uses default configurations and remains online for years. Even a slow adaptive worm could be difficult to remove if it has no central command server.
Anthropic projects that small models may reach a relevant capability threshold by the end of 2026, but that is an extrapolation, not an observed outcome. The research supports preparation rather than panic.
Vendors and network operators should disable exposed debugging interfaces, segment IoT devices, monitor unusual peer-to-peer traffic and plan for large model downloads. The investment implication is indirect: AI expands the attack surface, while basic device hygiene still determines whether a theoretical capability becomes an operational incident.