Anthropic Warns AI Could Self-Replicate on Smart Devices Soon

Your smart TV, refrigerator or security camera could one day become part of a network of self-replicating artificial intelligence systems. Anthropic has warned that smaller AI models may be only months away from gaining the ability to copy themselves across internet-connected devices without relying on traditional data centres.

An AI self-replicator is a system capable of creating copies of itself without direct human involvement. According to Anthropic, such systems could be difficult to contain once they spread across multiple devices because "there’s no centralised point to shut off," unlike a conventional data centre.

“Like coding agents in 2024, we’re seeing signs of life for AI worms,” Anthropic said. “We don’t yet have a full rogue replicator, but agents are becoming increasingly capable of completing the individual steps in the replication chain.”

Small AI Models Raise New Security Concerns

Anthropic's research examined weight-based replicators, which run AI models directly on compromised hardware. The company identified home and small-business IoT devices as potential targets because many remain connected to the internet and have historically offered limited security protections.

To assess whether AI models could operate on affordable hardware, Anthropic tested Gemma4-26B-A4B on 15 Android TV boxes costing around $55 (roughly Rs 5,200) each. The model ran at 0.39 tokens per second. A separate test found that a nine-billion-parameter model could run inference on eight of the nine hardware types examined.

The researchers found that models with 30 billion parameters or fewer can already complete many individual tasks involved in self-replication, although smaller models still struggle to complete the entire process reliably. Anthropic also said models such as Zai's GLM-5.2 were capable of end-to-end replication. If smaller models continue improving at their historical pace, they could reach similar capabilities by the end of 2026.

Smart Devices Could Form Large AI Networks

Anthropic's findings suggest that self-replicating AI would spread gradually rather than instantly. Setting up Gemma4-26B-A4B for inference took around 71 to 150 hours, while Qwen3.5-9B required 28 to 51 hours.

However, a large number of compromised devices could collectively support substantial AI computing capacity. Potentially affected devices include smart ovens, refrigerators, robotic vacuum cleaners, lawnmowers, security cameras and video-conferencing equipment.

The research also found that smaller models improved when they inherited notes about previous attempts. Qwen3.5-9B's inference setup success rate increased from 45 per cent in the first generation to 84 per cent in later generations.

Anthropic warned that manufacturers and security teams should strengthen protections for connected devices before such threats become more practical.

“We should aim to harden our systems to make it difficult for such replicators to spread, but also prepare to deal with a world where they are here to stay,” the company said.

While previous incidents involving rogue AI agents have occurred on company servers, the possibility of AI spreading across everyday household devices presents a different security challenge. If smaller models become capable of reliably reproducing themselves, containing them could become significantly more complicated.

ABOUT THE AUTHOR

Kahekashan

[email protected]

Kahekashan is a passionate technophile with a keen eye for cutting-edge gadgets, emerging technologies, and everything in the digital realm. Raised in a Defence family with strong values and a background in literature, she has consistently pursued excellence in every endeavour. Her last full-time assignment involved content writing with the Indian School of Business.

Read More