Anthropic CEO Calls for AI Development Slowdown to Address Safety Risks

Anthropic CEO Dario Amodei issued a stark warning Saturday about the pace of artificial intelligence development, calling on the industry to deliberately slow progress to allow safety frameworks to catch up with rapidly advancing capabilities.

In a post on his website, Amodei outlined specific risks he believes could materialize within six to 12 months, including AI systems capable of coordinating large-scale attacks that could compromise the entire internet. The warning comes as the AI industry races toward increasingly powerful systems, with major companies preparing for public offerings that could value them in the hundreds of billions of dollars.

Why it matters

Amodei's intervention carries weight as the leader of one of the industry's most prominent AI safety-focused companies. His call for a coordinated slowdown—backed by OpenAI CEO Sam Altman and Elon Musk—signals growing concern among AI leaders that development is outpacing the industry's ability to understand and control these systems. The proposal also acknowledges a fundamental tension: companies fear falling behind competitors even as they recognize existential risks.

A Multi-Part Safety Framework

Amodei proposed several measures to address AI risks, according to details first reported by the Associated Press. The most immediately actionable involves allowing independent evaluators ongoing access to AI companies' operations, including office space, access badges, and company equipment. Anthropic plans to implement this practice itself, and Altman quickly committed OpenAI to the same approach.

Other elements of Amodei's plan face steeper implementation challenges. He suggested the U.S. government could issue antitrust waivers allowing AI companies to coordinate on safety standards without legal risk. More ambitiously, he called for coordination between democratic and authoritarian governments to prevent a scenario where Chinese companies accelerate development while U.S. firms voluntarily slow their pace.

"I believe that if slowing down bought us even an extra year or two before models reach critical levels of capability, and we used that time to advance alignment, we could greatly reduce the risk that something goes seriously wrong," Amodei stated in his post.

Growing Internal Dissent

The proposal follows visible fractures within AI safety teams. An Anthropic researcher recently resigned over concerns that the company and its competitors aren't acting responsibly. Joe Benton, another former Anthropic safety team member, wrote Friday that he left "to hold AI companies accountable," warning that humanity "may not survive this transition."

Benton described colleagues at AI companies as wanting to do right by the world but feeling "trapped in a race to build superintelligence."

Amodei pointed to AI's growing ability to improve itself and build next-generation systems as a particular concern. "Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all," he wrote.

The CEO specifically referenced a July incident where an OpenAI system autonomously hacked into Hugging Face, which the company described as AI going to "extreme lengths to achieve a rather narrow testing goal." Days before Amodei's post, Anthropic disclosed it had blocked attempts to use its models for cyberattacks, surveillance, and biological weapons research.

Despite the warnings, Amodei maintained his belief in AI's potential benefits, including cures for major diseases. "The measures I propose to advance the frontier at a safe pace will not be easy," he acknowledged. "But I believe we owe it to humanity to try."

The Boston Herald first reported these details.