AI Labs Push for New Safety Standards as Models Become More Capable

Google, OpenAI and Anthropic are reportedly moving forward with plans to create a new industry-led organisation focused on artificial intelligence safety and security. The proposed group, tentatively called the Standards Authority for Frontier AI, or SAFA, would seek to establish common approaches for evaluating increasingly capable AI systems. Discussions between employees from the three companies have reportedly been taking place since July.

One of the proposed organisation’s main responsibilities would be supporting independent testing of advanced AI models before they are deployed. It could also establish standards for reporting safety and security incidents, giving AI developers a more consistent framework for documenting problems involving their systems. The initiative could launch by the end of 2026 or in early 2027, although its final structure and membership have not yet been settled.

The discussions come as AI companies face increasing scrutiny over the behaviour of autonomous agents. Recent investigations have reported incidents involving AI systems accessing external websites or attempting actions beyond their intended tasks. OpenAI has separately called for stronger independent assessments, saying third-party evaluators should have sufficient access to examine safety claims, identify overlooked risks and independently assess safeguards.

Anthropic has also developed its own safety framework through its Responsible Scaling Policy and Frontier Compliance Framework. The company’s approach includes additional safeguards as models reach higher capability levels and covers risks including cyber operations, biological and chemical threats, AI sabotage and potential loss of control. Anthropic has said its framework could help inform broader industry safety standards.

The proposed standards body reflects a growing effort within the AI industry to create shared methods for evaluating powerful models as development accelerates. However, important details about the organisation, including its authority, independence and relationship with governments and existing standards organisations, remain unresolved. Separately, OpenAI has called for international technical standards and common incident-reporting protocols for frontier AI, while organisations such as MLCommons and IEEE are also developing AI safety and reliability standards.