Anthropic warns that future is soon, but the threat is less novel than it seems.

Anthropic's Frontier Red Team recently published research suggesting that small AI models may be only months away from replicating and spreading across insecure devices, creating slow but hard-to-kill autonomous botnets. AI companies are no strangers to playing up the world-ending potential of their products as a form of marketing, but in this instance, experts Engadget spoke to broadly agree small AI models could become capable of self-replicating in the near future. However, the actual threat they pose may be less novel than it appears.

"It's maybe a little earlier than I would have said. I would have predicted something more like summer of next year as the time I'd be really worried about this," says Chris Canal, the CEO of EquiStamp, a company that works with AI labs, safety organizations and governments to verify and evaluate models. "Anthropic very likely has employees with information about the frontier of research that I don't have access to. So they very well could be making better predictions." Canal attributes the recent pace of model advancement to the fact that employees at Anthropic and other frontier labs are using their own AI tools to do machine learning research.

"I think [self-replicating models] are in the future. I think they're real. I think we'll see them," says Michael Bailey, a professor at the School of Cybersecurity and Privacy at Georgia Tech.

Source link