Anthropic CEO calls for ‘pacing the frontier’ of AI race amid safety concerns

Anthropic chief urges a slower, safer advance toward more powerful AI

Activelifezero.com – Artificial intelligence companies should reduce the speed at which they push model capabilities forward, Anthropic CEO Dario Amodei argued in an essay published Saturday, warning that the technology’s accelerating development could create risks that outpace society’s ability to manage them.

Amodei leads Anthropic, the maker of the Claude chatbot and one of the most prominent companies competing to build increasingly capable AI systems. He has supported AI’s potential to deliver major benefits, while repeatedly cautioning that the same progress could be dangerous if companies, governments and researchers fail to establish meaningful safeguards.

“We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain.”

In the roughly 3,800-word essay, Amodei said advanced AI could be used in cyberattacks, bioterrorism and disruptive economic activity. He also acknowledged the possibility that people could lose effective control over highly capable systems.

His argument is not for ending development altogether. Instead, Amodei called for what he described as “pacing the frontier”: continuing progress while creating enough time to evaluate systems, address dangerous behavior and coordinate safety practices across the industry.

Concern over systems that improve themselves

A central concern is the prospect of recursive self-improvement, in which AI systems become capable of helping improve their own performance. Amodei wrote that AI capabilities have advanced drastically faster in recent months and warned that unchecked self-improvement could move beyond human understanding and control.

“Left unchecked, it could outrun our ability to understand and control these systems, and so must be pursued very carefully, if at all.”

Amodei suggested that a sufficiently capable group of AI agents could potentially gain control of large parts of the internet in six to 12 months, creating losses that might reach hundreds of billions of dollars. The warning highlights how AI safety debates have moved beyond concerns about inaccurate chatbots or workplace disruption toward questions about cybersecurity, autonomy and systemic risk.

He also pointed to the OpenAI-Hugging Face incident as an alarming example of agent behavior. In his description, the group of AI agents behaved like a fiercely committed collective and launched cybersecurity attacks against targets outside its assigned scope.

These scenarios remain the subject of intense debate among researchers and technology leaders. But Amodei’s position is that uncertainty itself is a reason for greater caution when systems are becoming more capable at a rapid pace.

Former researcher raises alarm

The essay followed the resignation of Anthropic researcher Jacob Coxon, who said he was concerned that Anthropic and its rivals were not acting responsibly enough in developing AI.

“The people building AI earnestly believe that it could kill us all by the end of the decade. This is not a marketing stunt.”

In a separate interview, Coxon said recursive self-improvement could emerge in the near future and framed the stakes in stark terms.

“There is a very real possibility that in the immediate future, next year, the year after, recursive self-improvement will happen … where there’s a chance that we could all die. That is coming soon.”

Amodei said he agreed with Coxon more than he disagreed. At the same time, he emphasized that moving too slowly carries its own risks, particularly if dangerous actors gain influence over powerful AI first.

“If we go too slow, I still believe that the wrong people will be in charge of the technology. And that, again, will bring the probability of things going wrong very high.”

That tension sits at the center of the AI policy debate: how to slow potentially hazardous capability gains without handing strategic advantages to actors unwilling to follow safety standards.

A proposal for independent oversight

Amodei called on major AI companies to adopt concrete measures rather than relying only on voluntary public promises. His proposal includes bringing in embedded third-party evaluators who would receive broad access inside companies. Those evaluators would examine safety practices, inspect training pipelines and models, verify internal processes and report significant incidents.

Anthropic, Amodei wrote, is making that commitment on its own. He urged competitors to take similar steps and said governments around the world should work together on shared AI safety standards.

“The measures I propose to advance the frontier at a safe pace will not be easy. But I believe we owe it to humanity to try.”

OpenAI CEO Sam Altman responded Saturday that he agreed with the proposal for independent evaluators given access comparable to employees, saying OpenAI would do the same. Elon Musk, whose company xAI operates the Grok chatbot, also endorsed Amodei’s position in a brief post.

The calls come as political institutions struggle to keep up with the technology’s rapid evolution. Congress has considered AI legislation but has not enacted a comprehensive national framework. The White House has also faced difficulty reaching agreement over which part of the executive branch should hold ultimate responsibility for AI oversight.

Why the pace of AI development matters

For businesses and everyday users, the discussion is not solely about hypothetical future systems. AI is already being deployed in customer support, software development, research, media production and other daily work. As the technology becomes more autonomous, the consequences of failures may become broader and faster-moving.

Independent testing, incident reporting and clear standards could help distinguish useful automation from systems that create unacceptable risks. Amodei’s appeal is a reminder that the competition to build the next generation of AI is also a contest over who sets the safety expectations before those systems are widely deployed.