Google, OpenAI, and Anthropic are promoting the establishment of an industry-led AI safety standards agency, tentatively named the “Frontier AI Standards Authority”. The goal is to launch in late 2026 or early 2027 and operate independently without direct government supervision. The agency plans to establish voluntary safety and security standards for cutting-edge AI models, support third parties to conduct tests before model deployment, standardize how AI companies report security incidents, and set qualification standards for independent auditors. The three companies are also discussing whether SAFA will directly conduct model safety and capability tests. The plan previously considered adopting a public-private partnership model with government participation, but related progress has stalled and is now shifting to industry self-regulation. In the future, SAFA will also decide which models meet the “cutting-edge model” threshold and be included in its standard system, but issues such as the scope of membership and the division of labor with existing regulatory and security agencies are still being discussed.

Google, OpenAI, and Anthropic are promoting the establishment of an industry-led AI safety standards agency, tentatively named the “Frontier AI Standards Authority”. The goal is to launch in late 2026 or early 2027 and operate independently without direct government supervision. The agency plans to establish voluntary safety and security standards for cutting-edge AI models, support third parties to conduct tests before model deployment, standardize how AI companies report security incidents, and set qualification standards for independent auditors. The three companies are also discussing whether SAFA will directly conduct model safety and capability tests. The plan previously considered adopting a public-private partnership model with government participation, but related progress has stalled and is now shifting to industry self-regulation. In the future, SAFA will also decide which models meet the “cutting-edge model” threshold and be included in its standard system, but issues such as the scope of membership and the division of labor with existing regulatory and security agencies are still being discussed.