Anthropic: AI models will learn to reproduce in household appliances in a matter of months

Small AI models may have only a few months left before they learn to replicate themselves by moving from one cheap connected device to another—from smart ovens to no-name TV set-top boxes— warned AI developer Anthropic. If they continue to develop as quickly as they have so far, they will reach the necessary level by the end of the year—and they won’t even need data centers: in an experiment, Anthropic managed to run a “surprisingly powerful” model on 15 set-top boxes costing $55 each.

Anthropic refers to a program that copies itself from one computing device to another along with an AI model and runs it directly on the compromised device as an “AI worm.” Google’s Gemma4-26B-A4B was running on the TV set-top boxes. Anthropic classifies it as a small model—that is, one with no more than 30 billion parameters: Gemma has 26 billion. GPT-3, which OpenAI introduced back in 2020, had 175 billion.

In laboratory tests, only the GLM-5.2-based worm was able to complete the entire chain—finding an uninfected device on the network, copying itself onto it, launching, and moving on—but it is too large for infected machines, so it ran on a separate server. Furthermore, accessing these machines in the test required no authentication, so Anthropic estimates that the result would not have been any better on a real network. “There is no fully-fledged replicator that has gotten out of control yet, but the agents are getting better at handling individual steps in the self-replication chain (and passing on their progress to new copies),” Anthropic noted.

A basket of cybersecurity stocks tracked by Goldman Sachs more than doubled in price between its April low and the 20th of September: Investors expect that threats posed by advanced AI models will benefit the industry, Bloomberg reported at the time. Shares of cybersecurity providers CrowdStrike, Palo Alto Networks, and Fortinet rose by more than 130% during this period.

According to Brian Essex of JPMorgan, AI could lead to more than $430 billion in additional cybersecurity costs over the next three years, and demand for protection will remain strong regardless of attempts to “slow down” the development of neural networks. At the same time, detecting an AI worm is entirely feasible, Anthropic emphasizes: if it begins to spread across smart devices, internet service providers’ monitoring systems will detect anomalous traffic. “We must strive to strengthen our systems to hinder the spread of such replicators, but also prepare for a world in which they will remain forever,” the AI giant concluded.

Botnets on cheap TV set-top boxes still exist today, albeit without AI, and investors have already fallen victim to them. In June, researchers from several companies linked one of these botnets, Popa—a network of infected devices used for ad fraud and account hijacking—to NetNut, which is operated by the Israeli company Alarum Technologies, according to the Krebs on Security website. On July 2, Reuters reported that Google had suspended NetNut’s accounts and services, and on the same day, Alarum’s stock on Nasdaq plummeted. Alarum itself disagrees with the researchers’ findings: the company calls them “inaccurate and erroneous,” and describes NetNut as a “legitimate commercial proxy network.”

When Anthropic CEO Dario Amodei called on the industry in September to slow down the development of cutting-edge models, warning that in 6–12 months a swarm of agents “might be capable of taking over the entire internet,” “Short Seller” investor Michael Burry viewed such statements as beneficial to the AI companies themselves: “An IPO needs hype and self-promotion.” In the company’s IPO prospectus, which Reuters reviewed, 80 of the 261 pages are devoted to risk factors. According to Bloomberg, Anthropic plans to go public in mid-November.

Last Sunday, October 11, hackers disabled refrigeration equipment in hundreds of French supermarkets. The attackers gained access to the cooling systems and activated the defrost mode, causing the temperature in refrigerators and freezers to rise to 50 degrees for about six hours. The stores had to throw away the spoiled meat, vegetables, and frozen foods.

It is not yet clear who is behind the hack. Bloomberg notes that hackers are increasingly seeking not just to steal data or extort money, but to interfere with the operation of systems that control equipment and technical processes. The agency points out that water supply systems in the U.S., as well as oil and gas tankers, have previously been targeted by similar attacks.

This article was AI-translated and verified by a human editor