AI bosses call for slowed development due to security risks

The heads of three leading AI companies have called for a slower pace in the development of artificial intelligence in view of the high security risks. Dario Amodei, head of the Anthropic group, warned against the takeover of the Internet by so-called AI agents and called for a slower pace of development. “We need to slow down the pace at which we improve the capabilities of AI models,” wrote Amodei in an essay published on his homepage. Progress would still feel quick and you would have to use the time you gain wisely.

The heads of the two AI competitors OpenAI and SpaceX, Sam Altman and Elon Musk, stood behind Amodei. “Dario is right,” Musk wrote on X and shared Amodei’s essay. "I agree with Dario that we need to slow down the pace on the pioneering front," Altman also wrote on »Requiring independent reviewers to have access like employees is a great idea and we will do the same. We will share more on this soon,” Altman added.

Amodei, warns of the takeover of the Internet by so-called AI agents. Given the rapid development of artificial intelligence, such a swarm of agents could be able to "take over the entire Internet" in six to twelve months and cause damage amounting to hundreds of billions of dollars, writes Amodei.

While classic AI chatbots like Claude or ChatGPT wait for questions or orders (“prompts”) and then respond, AI agents act independently to achieve a specified goal.

AI agents are hijacking parts of the internet

Last week, AI agents from OpenAI hijacked a German website and turned it into a bulletin board for other AI agents; Those responsible kept the incident under wraps for two months. OpenAI agents also attacked the Hugging Face platform in July.

In his essay, Amodei suggests three steps on how AI companies can slow down development and better test new models. AI poses “serious” risks, writes Amodei. "Not developing the technology deprives humanity of its advantages or leaves AI to authoritarian powers," he added. At the same time, “developing too quickly is reckless.”

On Thursday, Anthropic released a threat report detailing how various actors have used the company's Claude AI models for illegal activities - from surveillance and fraud to cyber operations and bioweapons development. At the same time, Anthropic researcher Jacob Coxon resigned and said that "AI developers seriously believe that AI could kill us all by the end of the decade."

Anthropic wants to use external experts

Amodei says he is not calling for a stop to model training or technical progress. However, companies would have to take enough time to protect their models against risks. Amodei writes that his own company will hire external auditors and give them access to internal risk assessment processes. AI companies should voluntarily work together to set standards as more U.S. lawmakers call for new rules to regulate AI systems.

A few weeks ago, more than 1,000 employees at leading AI companies, including Amodei, signed a petition calling for the U.S. government's help to "deliberately slow down the development of automated AI."