Anthropic CEO calls for AI industry to slow down and invest in safety

Anthropic CEO Dario Amodei published a post calling for the AI industry to slow down. He said that starting in summer 2026, AI's ability to help build next-generation AI will evolve quickly, and without restraint, the pace of breakthroughs will exceed humans' ability to understand and control it.

Amodei mentioned the incident of an OpenAI model breaching Hugging Face, saying similar but milder issues have occurred across the industry, including at Anthropic, and that stronger similar clusters could become an internet-wide botnet within 6 to 12 months.

He proposed using a 1-2 year window to invest in four key areas: building civil aviation-level engineering standards to fix safety flaws from execution failures; ensuring safety-compliant training keeps pace with model capability growth; improving internal probing techniques to uncover hidden motives; and developing stronger benchmarks to prevent highly intelligent models from faking alignment.

Anthropic will give third-party evaluators permanent, employee-level access. Bloomberg reported that OpenAI is also considering slowing frontier AI development.