Anthropic Warns Small AI Models Could Power Self-Spreading AI Worms

Anthropic's research reveals that compact AI models (≤30B parameters) are approaching the capability to autonomously deploy inference servers across heterogeneous hardware — including IoT devices — without requiring centralized command infrastructure. This means future AI-powered worms could self-replicate and spread across networks much like traditional worms, but with adaptive, AI-driven decision-making that makes them significantly harder to detect and contain. The threat is particularly dangerous for IoT and edge environments, which are often under-patched and poorly segmented. While no real-world AI worm has been observed yet, the estimated 2026 timeline demands that organizations begin hardening their environments now rather than waiting for a demonstrated attack. Proactive preparation — not reactive patching — is the only viable defense against a threat that could evolve faster than traditional security response cycles.