IA · 14 October 2026 · 4 min read
Small Models and Millions of Smart Devices: Anthropic Warns of Self-Replicating AI Worms
In brief: A newly published experimental study by Anthropic reveals an emerging cyber threat: combining lightweight open-weight models with distributed computing techniques enables self-replicating AI worms on consumer IoT hardware. By running shards across low-cost devices like smart TV boxes and connected appliances, these autonomous agents can dynamically generate exploits and adapt to unfamiliar architectures, making traditional centralized takedowns ineffective.
by Team Mocchi's
For years, the public conversation surrounding artificial intelligence safety has remained anchored to frontier models: monolithic neural networks with hundreds of billions or trillions of parameters, tightly guarded inside hyperscaler data centers behind robust physical and network perimeters. A new experimental study published by Anthropic turns that premise on its head, pointing instead to the opposite extreme of the computing ecosystem: small open-weight models and the vast, unmanaged ocean of Internet of Things devices.
The paper outlines what Anthropic calls «small weight-based replicators»: an emerging category of autonomous malware where compact language models serve as computer worms. Capable of inhabiting smart TVs, routers, security cameras, and connected appliances, these systems could spread across local networks without relying on any centralized command-and-control infrastructure.
Android TV clusters and distributed model sharding
The technical backbone of the experiment relies on sharding and model quantization. By compressing model weights into lower-precision formats and distributing the compute load across peer nodes, researchers showed that hardware never intended for machine learning can still run modern networks. As reported by Ars Technica, Anthropic researchers linked fifteen budget off-brand Android TV boxes—each featuring just 4 gigabytes of RAM—and successfully achieved inference with the 26-billion-parameter Gemma 4.
Individually, none of those devices possessed the memory bandwidth or capacity to load such an architecture. Yet, operating cooperatively over a local network, they successfully processed prompts. Furthermore, researchers found that compact models such as Qwen 3.5-9B could autonomously inspect host hardware environments, correctly configuring and compiling runtimes across eight out of nine unfamiliar device setups without manual intervention.
Moving past traditional botnets to roaming weights
This architecture represents a fundamental departure from legacy threats like the 2016 Mirai botnet. Conventional malware relies on static, precompiled binaries engineered for specific instruction sets (such as ARM or MIPS), paired with remote command servers to issue tasks. If law enforcement or security researchers seize those command-and-control domains, the botnet collapses.
By contrast, an AI-augmented worm carries its reasoning engine with it: the model weights become an active component of the payload. Upon compromising an unfamiliar device, the agent can analyze the environment in real time, write custom scripts to exploit local vulnerabilities, and orchestrate the transfer of remaining model shards from peer devices or public repositories. According to The Verge, in simulated tests the GLM-5.2 model successfully executed the complete end-to-end chain of discovery, exploitation, and self-replication across a heterogeneous virtual Android network.
Anthropic describes the threat profile not as a lightning-fast kinetic strike, but rather as a «digital slime mold»: a slow, distributed infection with doubling times measured in days, yet exceptionally resilient and virtually impossible to sever through single-point mitigation.
Theoretical concern versus real-world risk
While the demonstration proves that each link in the operational chain is technically viable, the researchers caution that fully functional wild replicators remain difficult to build today. Latency over consumer wireless networks and resource exhaustion on bare-bones microcontrollers still pose major engineering hurdles to unconstrained propagation.
Nevertheless, the trajectory is clear. Small models are advancing in reasoning ability faster than frontier models are growing in size. Combined with the endemic lack of security updates and lifecycle management across consumer IoT hardware, the physical world offers fertile ground for autonomous software agents designed to survive in the wild.
Mocchi's take
Anthropic's findings validate a trend we have been tracking closely: radical model compression is driving machine intelligence deep into edge computing territory. For businesses managing operational technology, smart facilities, or distributed hardware fleets, this transition requires a refreshed defensive posture. We can no longer assume that resource-constrained devices are immune to complex software abuse simply because of their small footprint; strict network micro-segmentation and vigilant monitoring of peer-to-peer anomalous traffic remain our strongest lines of defense.