Security experts push back on Anthropic's warning about AI worms in TVs

Anthropic warned last week that AI models may be able to copy themselves without a datacenter, after running an AI model on the computing power of 15 smart TVs. Many cybersecurity experts say the warning misleads. They point out that self-copying malware is decades old, and that the hard part was never shown.

What Anthropic said

The company posted on X and on its website: "AI models may be capable of self-replicating without access to a datacenter." It also warned that other internet-connected devices could be used to run large language models, the kind of AI behind chatbots. The post set off anxious coverage. The Guardian even asked if there will soon be "AI in your oven?"

Why experts push back

Juan Andrés Guerrero-Saade, an adjunct professor at Johns Hopkins and a fellow at the security company SentinelOne, criticized the post in a long thread on X. He said "cyber research" is being used as a tool for policy arguments and regulatory capture. Researcher Anthony Jones told Straight Arrow that AI companies are not experts in hacking.

Jones also said Anthropic skipped the hardest step: finding exposed devices. "Once you're on the devices, 90% of the work's done," he said. In his view, the company ran a slow, weak AI model on cheap TVs and turned that into a scary headline.

An old problem

A worm is malware that infects one device, then reaches out and infects others. Jones says worms date to the 1970s. The Morris Worm in 1988 slowed almost 10% of the 60,000 computers that existed then. Its creator was the first person convicted under the 1986 Computer Fraud and Abuse Act. The Mirai botnet has been spreading for 10 years.

A botnet is a group of hacked devices used together, for example to knock a website offline. Smart devices suit this because they are often on all day, there are billions of them, many keep default passwords, and no rules govern their security. Jones said running AI would make a botnet slower, since it eats memory. He added that a smart fridge could already be part of one, and an owner might never know.

What to do

The practical lesson is simple. Change the default passwords on smart TVs, fridges, cameras and similar devices. Jones noted that defenders are "really pretty good at shutting them down now," but weakly protected devices remain the long-known weak spot.

What is not known

Jones said Anthropic's research did not show a botnet actually being created. The company concluded there are "signs of life" for self-replicating AI worms. How Anthropic answers the criticism in full is not clear from the available reporting.