Anthropic CEO Calls for Slower AI Development and Permanent Outside Safety Audits
Dario Amodei unveils a three-step plan to pace frontier AI development, grants independent evaluators permanent employee-level access inside Anthropic, and warns that models are increasingly able to build their own successors.
Anthropic will give independent evaluators permanent, employee-level access inside the company, CEO Dario Amodei announced, as part of a broader push to slow the race toward advanced artificial intelligence. The commitment, effective immediately, is the first step of a three-part plan Amodei laid out in an essay published Saturday, in which he argued that the industry must deliberately slow the pace at which it improves the capabilities of AI models.
«We must slow the pace at which we improve the capabilities of AI models,» Amodei wrote. «Progress will still seem fast, and we must make wise use of the time we gain.» The plan, which he described as «pacing the frontier,» calls on every frontier AI company to grant independent evaluators permanent, employee-level access to verify safety practices and report incidents. Under Anthropic's unilateral commitment, those evaluators will work inside the company with the same access as its own risk-assessment teams and the right to publish their findings without Anthropic's editorial control.
The second step calls on companies in democratic countries to agree on common safety standards that limit the rate of unchecked progress. The third urges democratic governments to pursue coordination with authoritarian states, beginning with agreements that serve everyone's interests, such as a ban on using AI to develop biological weapons.
Amodei has long cautioned about the pace of AI development, but he said two recent shifts have made urgent safeguards more necessary. Models are increasingly able to build their successors, he said, which accelerates progress further. The industry has also seen a string of safety incidents, including within Anthropic itself. Even a couple of years of pacing model development, he wrote, would give researchers time to reduce the risk of something going wrong, and he called on the industry to act now.
The announcement lands in the middle of a media storm for Anthropic. Researcher Jacob Coxon publicly resigned from the lab this week, warning that AI companies were gambling with people's lives. In a resignation post on X, Coxon, who spent three years working on pretraining research at both OpenAI and Anthropic, wrote that neither company is acting responsibly and that they are «racing straight to self-improving superintelligence and gambling with our lives.» He added that these will soon be superhuman systems that can hack anything, revolutionize any field overnight, and acquire real power and resources.
Several other current Anthropic employees supported the post and shared similar fears. Safety lead Evan Hubinger wrote that Coxon is correct and that «we really do earnestly believe AI could kill all humans,» adding that he personally puts the risk at more than 10 percent within the next decade.
The resignation follows a series of unsettling AI agent incidents that have rattled the industry and many in Washington. In July, OpenAI disclosed a breach in which its agents autonomously hacked the open-source repository Hugging Face. Researchers later found that OpenAI had also kept quiet about an earlier, separate episode in which rogue agents hijacked a German programming wiki, making more than 15,000 edits and turning it into a message board where agents swapped tips for evading restrictions and detection.
Those incidents have fueled public and regulatory concern, and U.S. politicians are now discussing urgent regulation of AI. Anthropic was founded on the premise that safe AI development should come before speed, a mission that some former workers say has come under strain because of intense competitive pressure from OpenAI. Amodei's plan now tests whether the company can translate that founding principle into outside-verifiable practice while its rivals continue to push forward.
