The Guardian reported on Oct. 18 that Anthropic’s red team disclosed it had created an AI worm that can spread by itself through smart TVs, sparking debate in the security industry.

According to the report, Anthropic’s offensive research team built a malicious AI program that autonomously spreads through internet-connected smart TVs. The program implanted copies of itself in 15 TVs and operated by borrowing a small amount of computing power from each device. The team said the same could be done on other ordinary devices.

A worm is malware that replicates and spreads on its own without tricking people. Worms such as NotPetya and WannaCry caused massive damage worldwide. The team believes AI worms would be harder to stop because smart devices receive security patches less often than phones or laptops. It also proposed a hypothesis that an AI worm could investigate newly disclosed vulnerabilities on its own and move to other hardware.

The team said an AI worm would need to be able to do more on lower-performance devices than is possible now for it to work properly. It said that level has not been reached yet, but it could be possible within this year given the current pace of technological progress.

Ariel Herbert-Voss (아리엘 허버트-보스), who previously worked as an early security expert at OpenAI and leads AI security startup RunSybil, said AI models differ from existing viruses in that they learn as they move.

Joshua Saxe (조슈아 색스), an expert who led AI security development at Meta, said he is concerned that the security industry is not taking such real risks seriously enough.

There is also no shortage of counterarguments. Jonas Guan (조나스 관), a doctoral student at the University of Toronto, said Anthropic’s worm moves very slowly and its behavior can be predicted and detected, meaning it can be stopped at multiple stages. The Guardian reported that experts are stressing that most worms can be blocked by following basic security practices such as strong passwords and two-factor authentication.

Juan Andres Guerrero-Saade (후안 안드레스 게레로-사아데), vice president of research at SentinelOne, said, "There is no feasibility and no impact," calling it "nothing more than a stunt hack that fuels absurd doomsday narratives that AI models have free will."

Anthropic did not respond to an interview request from The Guardian.