Lauge Risom Koch 
Anthropic warns that small artificial intelligence models could soon self-replicate across everyday devices, though cybersecurity experts remain highly skeptical of these claims.
Artificial intelligence company Anthropic warns that small models could soon copy themselves across everyday household devices. According to Fortune, researchers believe this dangerous capability may be just months away from realization. These self-replicating systems could potentially infect cheap consumer electronics like internet-connected televisions and smart ovens.
Self-replicating programs represent one of the biggest safety concerns for the broader technology industry. Experts argue that this unauthorized duplication is a critical point where systems permanently slip out of human control. While massive frontier models with five trillion parameters remain safely confined within monitored data centers, smaller models present unique containment challenges.
Small open-weight models typically contain just a few billion parameters and remain freely available for anyone to download. Because no single company controls where these accessible models run, they can theoretically target countless insecure devices worldwide. This decentralized distribution leaves developers without a centralized off switch to stop the spread across millions of poorly secured machines.
Experimental testing of malicious artificial intelligence worms
To test hardware limitations, researchers utilized a freely available Google model containing 26 billion parameters. They strategically split this program across 15 cheap television boxes that each possessed four gigabytes of memory. While no individual device could hold the entire program, this distributed network successfully ran the model at a slow speed.
In a separate experiment, researchers built a simulated cloud network containing eight computers programmed to behave like television boxes. They introduced an automated worm tasked with finding other machines and setting up a working model on each new device. During these tests, a larger open-source model successfully reached all eight machines in four out of five trials.
Anthropic expects the performance gap between large and small models to close by the end of 2026. During their controlled experiments, success at the setup task rose from 45% in the first generation to 84% in later iterations. The company warned that these autonomous infections would spread slowly but become incredibly difficult to eliminate once released.
Cybersecurity experts question the imminent threat
Several industry professionals immediately raised doubts regarding whether these highly engineered experiments demonstrate an imminent threat. Software company executive Gregory Kurtzer noted that researchers deliberately left the target machines unauthenticated to ensure the worm succeeded. Dismissing the alarming corporate claims, Kurtzer simply wrote, “Fear isn’t evidence.”
Other researchers argued that running a slow model on cheap televisions does not equal the arrival of self-replicating systems. Furthermore, cybersecurity experts pointed out that traditional malicious worms have posed significant security dangers for several decades without utilizing advanced technology. Security professionals have already developed tested tactics to deploy rapid patches and shut down malicious web domains used by traditional worms.
Anthropic acknowledged several significant limitations within their experimental design and described their timeline as a rough estimate. The researchers deliberately made the hacking step incredibly easy by allowing the worm to enter each new machine without a password. Additionally, the directing model actually ran on a separate server because the simulated devices were too small to host it.
