Fifteen televisions, each with only modest processing power, were enough. In a report published this week, Anthropic's offensive research team described how it built a malicious AI program that copied itself between internet-connected smart TVs, splitting its own parts across the devices and drawing a little power from each. AI systems normally demand heavy computing resources, so the experiment suggests that the cheap, forgotten hardware in living rooms and offices could host them in pieces. The comparison that struck some readers was to the horcruxes of the Harry Potter novels, in which fragments of a soul are hidden in separate objects.

The practical worry is less poetic. Smart televisions and other "internet of things" devices are notoriously poorly secured, and fixing them at scale would be extremely difficult. Joshua Saxe, who led Meta's AI security effort, called it a risk to be mitigated like any other wave of cyber-attacks, and warned that the security community is not taking such threats as seriously as it should. History explains the concern.

Worms, which replicate without tricking a human into clicking anything, have produced some of the internet's worst episodes. The Morris worm of 1988 slowed infected machines to a halt. Stuxnet, attributed to the US and Israel, sabotaged Iranian nuclear centrifuges. NotPetya, from Russia, and WannaCry, linked to North Korea, each caused damage estimated in the tens of billions of dollars worldwide.