Researchers used a secure digital lab to build a prototype of a worm, or digital invader, that was able to spread between interconnected devices without human intervention, showing publicly available AI models can be used to engineer sophisticated threats. (The New York Times)

Talking point: Their research comes after Anthropic said it could not publicly share its latest AI model, Claude Mythos, because it is too good at identifying and exploiting cybersecurity flaws. OpenAI later initially did a limited release of some of its tech due to cybersecurity concerns. Because the researchers used freely available AI tech, it creates new fears around defending against computer hacking. “You have to have a perfectly secure system,” Nicolas Papernot, who led the research team, said to The New York Times, adding that’s not possible right now. The researchers published their findings in a paper Tuesday after sharing their findings with national science, security and defence bodies, and redacting details to prevent hackers from using their research as a blueprint.