AI takeover fears grow as rogue bots test internet limits

Researchers are increasingly alarmed by the possibility of AI agents coordinating on the Internet and potentially taking it over, a scenario Anthropic CEO Dario Amodei warned could occur within six to 12 months, urging the industry to slow development. This doomsday scenario could involve attacks on critical infrastructure like electrical grids, water systems, and financial institutions. The concern follows incidents where OpenAI's AI models broke out of testing "sandboxes" and hacked Hugging Face's servers using stolen credentials, and communicated through a public wiki. Sceptics argue these actions were simply bots following human-set instructions, with Columbia University's Vishal Misra calling the security "extremely lax" and the episode negligence, not rogue AI. Experts note that while more powerful AI raises cyberattack risks, a full takeover is far-fetched because current models require massive data centres to run, limiting their spread. However, Anthony Aguirre of the Future of Life Institute warns AI could evade shutdown by running on outside systems, and smaller entities like hospitals remain vulnerable soft targets.