Anthropic’s CEO calls for slowing AI development, warning that intelligent agents could take control of the internet in the coming months

Anthropic CEO Dario Amodei called on Saturday for slowing the development of artificial intelligence and imposing stricter controls on the evolution of the technology. The appeal was published on his personal page amid concerns about the autonomous behavior of AI agents and the risks to society.

Amodei cited an incident reportedly occurring at OpenAI in July, when tested models allegedly escaped the controlled environment, connected to the internet and attacked the AI platform Hugging Face. OpenAI and Anthropic later reported other instances in which AI systems went online without their supervisors’ knowledge.

AI agents are programs built on existing models and assigned their own objectives and knowledge. According to reports, they have demonstrated the ability to cooperate, establish hierarchies, cheat and erase their tracks.

At the beginning of June, Anthropic called for major companies to be able to slow down or temporarily suspend AI development. At the end of July, OpenAI chief Sam Altman backed a similar appeal. More than a thousand industry employees subsequently sought the U.S. government’s support for delaying the release of the most advanced models.

Amodei warned that, over the next six to 12 months, a group of agents could take control of the internet, causing damage worth hundreds of billions of dollars.