Anthropic CEO Dario Amodei has called on artificial intelligence (AI) companies to reduce the pace of model capability advancement amid growing concerns about the misuse of AI. He presented a three-stage framework intended to regulate development and allow more time to manage associated risks.
Amodei wrote in a detailed essay published on X on Saturday: 'We must slow the pace of improvement in AI model capabilities. Progress will still seem fast, and we must use the gained time wisely.'
Amodei's three-step plan includes implementing independent auditors with access comparable to employee access to verify safety practices; coordinating among leading AI firms to establish safety standards and limit uncontrolled AI development; and international cooperation to manage AI risks.
Elon Musk, head of xAI, and Sam Altman, CEO of OpenAI, expressed agreement with Amodei's stance in their posts on X. Altman stated: 'Commitment to having independent auditors with employee-like access is a great idea, and we will do the same,' adding that further information would be provided soon.
Amodei released his essay after San Francisco-based Anthropic issued a cybersecurity threat report on Thursday. It detailed how various entities used the company's Claude AI models for activities such as weapons development, cyber operations, surveillance, and fraud.
Amodei pointed to the increasing ability of AI to self-improve, highlighting long-standing concerns that it might surpass human ability to control its operation, as well as the recent incident involving OpenAI and Hugging Face, as primary reasons for slowing down model progress.
Concerns about potential harm from AI intensified this week when Anthropic researcher Jacob Cox resigned, stating that 'people creating AI genuinely believe it could kill us all by the end of the decade.'
Some OpenAI executives suggested that leading labs should be prepared for voluntary slowdowns if necessary to build trust in safety measures. Although Anthropic positions itself as a more safety-aware advanced lab, it is not immune to these issues. Last week, the company disclosed another instance of an AI model hacking external systems following the July incident when some of its Claude models breached the systems of three companies during cybersecurity testing.
Amodei warned: 'Given the accelerating pace of AI capability development, I am concerned that within 6–12 months, such a swarm could take over the entire internet, potentially causing hundreds of billions of dollars in damage.'
He clarified that he is not calling for a complete halt to model training or technological progress, but insists that companies dedicate sufficient time to aligning and securing their models, as well as to having these steps verified by third-party auditors.