The three companies most responsible for the AI arms race are now trying to negotiate a ceasefire. Anthropic, OpenAI, and Google DeepMind have been quietly discussing the formation of an industry-led safety standards body since July 2026, a conversation that burst into public view when Anthropic CEO Dario Amodei published an essay on September 12 calling for a voluntary slowdown in AI development.
The three-step plan
Amodei’s essay lays out a surprisingly concrete framework. His proposal rests on three pillars.
First: independent evaluators who would receive employee-level access to frontier AI models. Not a quick demo or a polished benchmark, but the kind of deep, sustained access that lets auditors actually stress-test what these systems can do.
Second: coordination among democratic nations to align on safety standards.
