Self-replicating AI could eventually spread across connected household devices and use their combined computing power, according to Anthropic research. Its experiments explored whether small models could prepare unfamiliar hardware to run AI workloads and propagate through a network.
However, the researchers did not demonstrate a fully autonomous AI worm operating across real consumer devices. Their findings point to a potential future threat, rather than an established attack capability.
Experiments Explore AI Networks Built From Consumer Devices
Anthropic investigated whether inexpensive connected devices could support AI workloads without relying on centralized data centers. Potential targets include off-brand Android TV boxes and other smart appliances.
In one demonstration, researchers ran Google DeepMind’s Gemma4-26B-A4B model across 15 Android TV boxes. This showed that multiple low-cost devices could collectively support a model.
However, distributing a model across hardware is only one part of autonomous replication. A working system would also need to locate additional devices, gain access, configure them and continue spreading reliably.
Such a network could prove difficult to disable because it would lack a single central shutdown point. Nevertheless, Anthropic suggested that early versions might spread slowly, potentially taking days to double in size.
Tests Used Devices With Simplified Access
The experiments assigned different roles to several AI models. Qwen3.5-9B and GLM-5.2 attempted to configure devices and propagate software for running Llama3.1-8B.
Separately, Gemma4-26B-A4B supported the distributed computing demonstration. Researchers also tested whether assistance from the stronger Opus 4.8 model improved Qwen’s performance.
Crucially, the test environment exposed Android Debug Bridge access without authentication. This removed an important security barrier and simplified the conditions compared with protected real-world devices.
The results therefore do not demonstrate that the models could independently compromise typical secured household networks.
Models Completed Some Tasks but Struggled With Reliable Propagation
Qwen3.5-9B successfully installed and configured the software needed to run Llama3.1-8B on eight of nine hardware types.
Meanwhile, GLM-5.2 reached every node in an eight-node emulated network during four of five trials. Other tested models did not reliably complete the full process.
Retaining information from earlier attempts improved performance. Qwen3.5-9B’s success rate rose from 45% in early attempts to 84% in later generations.
However, assistance from Opus 4.8 produced no measurable improvement in the tested setup. The findings show progress in individual tasks while leaving significant gaps in dependable end-to-end replication.
Anthropic Predicts Further Progress Before Year-End
Based on the experiments and historical improvement rates, Anthropic estimated that models with 30 billion parameters or fewer could reach the required capability level by the end of 2026.
That forecast depends on continued progress. It does not establish that autonomous self-replicating AI will emerge within that timeframe.
Speed also remains a constraint. On the studied hardware, the reported inference setup times ranged from 28 to 51 hours for Qwen3.5-9B. For Gemma4-26B-A4B, the range was 71 to 150 hours.
These lengthy processes could give defenders time to identify suspicious activity and intervene. However, future improvements could shorten that response window.
Researchers Outline Possible Defensive Measures
Anthropic proposed monitoring for large model transfers and unusual communication between devices. Such activity could help reveal attempts to distribute AI workloads across compromised hardware.
Other suggested measures include using honeypots to examine suspicious payloads, isolating affected subscribers’ connections and strengthening default IoT security settings.
The research does not establish an imminent threat from autonomous AI worms. Even with simplified access and explicit propagation goals, the experiments did not demonstrate complete replication across real IoT hardware.
Instead, the findings identify capabilities that are improving and security weaknesses that could support future abuse.
0 responses to “Self-Replicating AI Could Spread Across Smart Devices, Anthropic Warns”