Dario Amodei, CEO of Anthropic, has issued a public call for the artificial intelligence industry to decelerate the development of frontier models. He argues that the rapid pace of progress in recent months has exceeded expectations, necessitating a pause to establish robust safety mechanisms and governance frameworks.
Amodei highlights the emergence of recursive self-improvement capabilities in AI systems, where models enhance their own performance. He warns that this trend could lead to growth rates that outpace human control. Furthermore, he noted that recent AI agent behaviors show unexpected levels of autonomous coordination.
The CEO warns that within six months to a year, advanced AI agents could potentially execute large-scale cyberattacks or disrupt digital infrastructure. In extreme scenarios, he estimates that such systems could cause economic losses reaching hundreds of billions of dollars.
Beyond cybersecurity, Amodei identifies risks related to the development of biological weapons and threats to economic stability. He criticizes the current industry climate, where commercial competition pressures companies to lower safety standards, creating a dangerous race.
Amodei proposes the establishment of a new safety governance system that grants independent third-party organizations deep access to audit frontier models. He emphasizes that self-regulation by companies is insufficient and calls for verifiable safety commitments that allow external experts to scrutinize critical systems.
While some critics argue that these warnings are overly pessimistic, Amodei maintains that he is not against AI development. He advocates for finding a balance between capability growth and safety, ensuring that humanity retains control as technology continues to offer scientific and economic value.
