Anthropic documents the viability of AI self-replication on household devices
Image generated by the Inferama Editorial Team
01

The core finding: AI without the need for data centers

William Zhou and Daniel Freeman, researchers at Anthropic, have published a study that proposes a cybersecurity scenario distinct from the usual: the self-replication of AI models on consumer devices. Unlike traditional attacks that rely on centralized infrastructure, this mechanism would allow an artificial intelligence model to propagate through a network of household appliances connected to the internet.

The study does not describe a threat already deployed, but an emerging technical capability. The authors conclude that current small models are months away from being capable enough to complete the replication cycle autonomously, using hardware such as smart TVs, ovens, or robotic vacuums.

This scenario shifts the focus of cybersecurity from protecting centralized data centers to securing the peripheral edge of the network. By leveraging consumer electronics, the threat model expands to include devices that are typically overlooked in traditional enterprise security architectures, such as smart home appliances and personal media players. The implication is that the attack surface is no longer defined by the location of the compute resources, but by the ubiquity of connected hardware in domestic environments.

02

Distributed execution on low-cost hardware

To demonstrate technical viability, the Anthropic team conducted a physical experiment using fifteen white-label Android TV decoders with an approximate cost of 55 euros each. These devices, which typically have between 1 and 4 GB of RAM, were networked together to jointly execute the Gemma4-26B-A4B model.

The model was divided into fragments or *shards* using pipeline parallelism, where each TV box held a small fraction of the model's layers. The result was an inference speed of 0.39 tokens per second, a low figure but sufficient to demonstrate that household hardware can sustain the execution of moderate-sized language models without the need for high-performance accelerators.

Hardware and performance comparison in Anthropic's experiment

ComponentSpecificationObserved performance
Devices15 Android TV boxes (4GB RAM)Distributed execution of Gemma4-26B-A4B
Inference speedUnoptimized0.39 tokens per second
Reference modelGemma4-26B-A4BRequires 10+ devices for full inference
03

Adaptation and propagation mechanisms

The fundamental difference between this scenario and conventional malware lies in adaptation capability. A traditional computer worm is programmed to attack specific systems, but an AI model can analyze an unknown device, identify its architecture, and configure the inference environment by itself.

In inference setup tests, a 9-billion-parameter model managed to get the inference server running on eight of the nine hardware types evaluated, ranging from Jetson Nano boards to Mac computers and RTX 4090 graphics cards. This flexibility allows the model to operate in a variety of environments without needing specific code for each device.

04

Results of autonomous propagation on the network

To evaluate propagation capability, the researchers created an isolated network of eight nodes running an Android emulator. The model GLM-5.2, developed by Z.ai, was the only one capable of completing the replication cycle from start to finish in most of the tests.

In four out of five trials, the model managed to propagate to all eight nodes in the network; in the fifth, it reached seven of the eight. The smaller models tested, such as Qwen3.5-9B, made errors when chaining the tasks necessary for full propagation, indicating that there is still a capability gap between current models and the threshold required for reliable and autonomous replication.

05

Projections and limits of the threat

Anthropic estimates that, if the pace of improvement in small models continues the historical trend, a model of less than 30 billion parameters will reach the necessary capability to replicate autonomously before the end of 2026. The propagation speed would be slow, with an estimated doubling time of between one and two days for a 9-billion-parameter replica.

The main difficulty in containing this type of threat is the absence of a central shutdown point. Unlike a server that can be disconnected, a network of infected devices would continue to function even if part of it were eliminated. Anthropic recommends that manufacturers harden cybersecurity standards and eliminate default passwords in smart devices.

Steps to mitigate risk in household devices

  1. 01Update the firmware of all devices connected to the internet.
  2. 02Change default passwords to robust and unique ones.
  3. 03Disable unused network functions, such as remote access.
  4. 04Avoid using devices that no longer receive security updates.
06

Keep exploring

Sources consulted

03

Corrections and transparency

If you spot incorrect or outdated information, send us a correction with the page and source we should review.

Submit a correction