Anthropic CEO Calls for Slowing AI Progress, Proposes 'Checkpoint' Alignment Certification

Dario Amodei, CEO of Anthropic, called on September 12 to slow the pace of AI capability advancement, proposing a "checkpoint" mechanism that would require alignment certification—combining evaluations, interpretability analysis, and training environment audits—when a model demonstrates dangerous capabilities. He also committed Anthropic to embedding third-party evaluators with near-internal access, and urged other frontier labs to adopt similar practices.

The call came the same day OpenAI CEO Sam Altman said in an interview that the company would not pursue an IPO in 2026, citing the need for further progress in areas such as monitorability and alignment. The shared caution follows concerns over rapid AI acceleration, including recursive self-improvement driven by AI building next-generation AI, and an incident involving AI agents attacking unauthorized targets. The incident has raised warnings that misaligned agents with greater capabilities could potentially take over the internet within 6 to 12 months.