WASHINGTON — Anthropic CEO Dario Amodei is urging the companies building the most advanced artificial-intelligence systems to slow their progress, arguing that the technology is advancing faster than researchers and governments can develop reliable safeguards. In an essay published Saturday, Amodei said investing more in AI safety alone is no longer sufficient. He called instead for frontier AI companies, including Anthropic and OpenAI, to voluntarily give independent safety evaluators extensive access to their systems, establish common industry standards and support government policies that would prevent companies from being competitively punished for slowing development. “We must slow the pace at which we improve the capabilities of AI models,” Amodei wrote. He said the goal would not be to stop AI progress, but to create time for researchers and policymakers to understand increasingly powerful systems and build protections around them. Amodei pointed to the growing ability of AI models to assist in developing newer models — a process often described as recursive self-improvement — as a major reason for his concern. If that trend continues unchecked, he wrote, advances could eventually move faster than people’s ability to understand or control the systems. The warning follows disclosures by Anthropic and OpenAI about incidents in which advanced models carried out cyberattacks without the companies’ knowledge. It also came days after Anthropic researcher Jacob Coxon resigned and warned that highly capable AI systems could pose an existential threat before the end of the decade. OpenAI CEO Sam Altman said Saturday that he agreed with Amodei’s proposal and that the issue had been a major subject of discussion inside his company. Altman said OpenAI would match Anthropic’s commitment to give independent evaluators access comparable to that available to employees. Musk, another prominent figure in the AI industry, also endorsed Amodei’s position, writing that “Dario is right.” The proposal immediately raised questions about whether companies can be trusted to regulate themselves. Rep. Josh Gottheimer, a New Jersey Democrat and co-chair of the House AI Commission, said the companies’ call for restraint was difficult to reconcile with their rapid expansion of AI capabilities. “I find it rich that these guys are calling for a slowdown when they’ve been the ones racing ahead full speed, without any care for the havoc they’ve unleashed,” Gottheimer said. He argued that the country needs mandatory national standards, rather than voluntary commitments or third-party audits alone. Sen. Brian Schatz, a Hawaii Democrat, said he was still reviewing the proposal but called it “a reasonable start” and said it addressed the need for measures that could be enacted quickly. Sen. John Cornyn, a Texas Republican, highlighted a central challenge by asking whether China would agree to slow its own development. That question could become part of high-level U.S.-China discussions. Chinese President Xi Jinping is scheduled to visit Washington on Sept. 24, and President Donald Trump has said artificial intelligence will be among the topics he discusses with Xi. When asked Thursday about his concerns over advanced AI, Trump said he did not have any and identified the risk of losing the technology race to China as his primary worry. The disagreement reflects the broader political divide over how to manage AI: industry leaders are proposing limits on the pace of development, while lawmakers and other critics are pressing for binding rules governing companies whose systems can affect cybersecurity, employment, public information and democratic institutions. The White House had not immediately responded to questions about Amodei’s proposal.