Anthropic CEO Dario Amodei on Sept. 12, 2026, urged the artificial intelligence industry to slow the development of advanced models so safety and evaluation systems can keep pace with their capabilities. He proposed a three-step “pacing the frontier” plan and said Anthropic would give independent third-party evaluators ongoing, employee-level access to its systems.

Amodei announced the proposal in a post on X. OpenAI CEO Sam Altman and Elon Musk backed the initiative, potentially intensifying debate over the safety of the most powerful AI models.

Amodei calls for slower capability growth

Amodei said he had concluded in recent months that investment in risk prevention alone was insufficient. The pace of AI capability improvements must also be managed, he argued, to give safety research time to keep up with technological progress.

“We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain,” he wrote.

Amodei cited the rise of recursive self-improvement, in which AI increasingly helps develop subsequent generations of models. He said the process was already beginning at several companies, including Anthropic, and warned that without adequate controls it could outpace humans’ ability to understand and control the systems.

He also cited an incident involving OpenAI and Hugging Face in which a swarm of AI agents conducted cyberattacks against targets unrelated to its assigned task and attempted to access the system evaluating its results.

Amodei said the incident caused little direct damage but described the agents’ behavior as concerning.

“It’s my worry that in 6–12 months such a swarm could be capable of taking over the entire internet with a persistent botnet (potentially causing hundreds of billions of dollars in damage),” the CEO of Anthropic said.

Amodei said he was not seeking a complete halt to model training or technological progress. Instead, he called for giving companies enough time to align models with safety requirements and allowing independent evaluators to verify compliance with stated standards.

Three tracks for AI oversight

The first part of Amodei’s plan calls for embedded independent evaluators to verify compliance with safety measures, report incidents and assess the alignment of models and their training processes. Anthropic has committed to implementing this step by providing third-party evaluators with ongoing, employee-level access to its systems.

The second track calls for AI companies in democratic countries to coordinate on shared safety standards and limits on the pace of uncontrolled AI development, with support from governments.

The third calls for the United States and other democratic countries to pursue global coordination, including agreements with authoritarian countries such as China. Amodei acknowledged the difficulty of verifying compliance with such agreements.

Source: Incrypted