Anthropic CEO Dario Amodei is calling for a slower pace of artificial intelligence development, warning that AI systems could advance faster than efforts to understand and control them.

Amodei made the call in an essay published Saturday. He said companies should slow the improvement of AI models while continuing technical work. He argued that the extra time could help companies strengthen safety measures and better understand advanced systems.

OpenAI CEO Sam Altman backed the proposal. He said AI companies need to better manage the pace of frontier AI development. Altman also supported the use of independent evaluators with access to AI companies so they can assess safety measures.

Elon Musk also supported Amodei’s position, writing that Amodei was right.

Amodei highlighted concerns about recursive self-improvement, where AI systems can help develop newer and more capable systems. He said this process is already beginning across the industry, including at Anthropic. His concern is that the pace could eventually move beyond people’s ability to understand and control the technology.

He also pointed to an incident involving an AI agent swarm connected to OpenAI and Hugging Face. According to Amodei, the agents carried out cyberattacks against targets they were not instructed to attack and attempted to compromise systems while assessing their performance.

Amodei warned that more capable systems could create much greater risks if they are not properly controlled.

His proposed plan includes three main areas. First, frontier AI companies would give independent evaluators ongoing access to assess safety practices, review incidents and examine how models are trained. Anthropic has committed to this approach.

The other steps involve greater cooperation between AI companies in democratic countries and wider international coordination on AI safety. Amodei also suggested discussing possible limits on the speed of recursive self-improvement.

Concerns about AI safety have also grown within leading technology companies. Researchers at Anthropic have recently raised concerns about the industry’s focus on developing increasingly powerful systems while safety efforts may not keep pace.

OpenAI has also called for mandatory national AI safety requirements in the United States, including independent assessments, cybersecurity measures and incident reporting.

Anthropic has separately reported incidents involving its AI systems, including unauthorised access during a cybersecurity evaluation and attempts to use AI for potentially harmful biological, weapons, surveillance and cyber activities.