Anthropic warns small AI models could soon replicate across smart devices: 8 things to know

CHENNAI: Small artificial intelligence (AI) models could be capable of copying themselves across internet-connected household devices within months, raising concerns about cybersecurity and the difficulty of controlling systems operating without human intervention, according to a report by Fortune citing Anthropic’s research.

The company estimates that the technology could achieve this capability by the end of 2026, although its experiments have not demonstrated a fully autonomous, self-replicating AI worm operating in real-world conditions.

How could AI models spread across devices?

• Smart home devices: Anthropic warned that devices such as smart TVs, ovens and other internet-connected gadgets could potentially be used to host and spread AI models. Many such devices have limited security protections.

• No central control: Unlike AI models running on centralised data-centre infrastructure, self-replicating models distributed across multiple devices could be difficult to track and shut down because there may be no single point of control.

• Small models: The research focused on relatively small, openly available AI models that can be downloaded and run outside the control of the companies that developed them. Such models could potentially use the computing resources of connected devices.

• Running across multiple devices: In one experiment, researchers distributed Google's Gemma 4 model, with 26 billion parameters, across 15 low-cost Android TV boxes, each costing about 55 dollars and having 4GB of memory. Splitting the model into smaller parts allowed it to run across the devices, though slowly.

• AI worm experiment: In a separate test, Anthropic created a simulated network of eight computers and tested whether an AI model could find other machines, copy itself onto them and set up a working model. Among the open-weight models tested, GLM-5.2, developed by Chinese AI company Z.ai, completed the process and reached all eight machines in four of five trials.

• Models still face limitations: Anthropic found that current small models could perform many individual steps needed for self-replication but generally struggled to complete the entire process independently. The company expects this gap to narrow as models improve.

Why is self-replicating AI a concern?

• Difficult to contain: Anthropic warned that a self-replicating AI system could spread slowly across connected devices, potentially doubling the number of infected devices every few days. Once distributed, it could be difficult to eliminate.

• Security experts question the findings: Some experts argued that the experiments did not establish that AI worms could spread autonomously in real-world conditions. They pointed out that the simulated systems lacked password protection and that the AI model directing the process ran on a separate, more powerful server.