OpenAI proposes global safety standards for self-improving AI

OpenAI proposed on September 21, 2026, a set of global technical standards to guide the safe development of frontier artificial intelligence. In a blog post reported by CNBC, the company advocated for international cooperation and recommended building on the work of AI safety institutes worldwide. The proposed standards focus on alignment research and managing the risks of automated AI researchers, including systems capable of recursive self-improvement (RSI).

Recursive self-improvement allows foundational models to refine themselves without human intervention, raising concerns among specialists about potential loss of control. OpenAI stated in its post that fully autonomous RSI is not occurring today and argued that developers should not pursue it until it can be done safely. The company referenced an attack on a Hugging Face agent as an example of risks that could worsen without safeguards, although that incident did not involve RSI.

The proposals follow public warnings across the artificial intelligence sector. Jacob Coxon, a researcher who previously worked at both OpenAI and Anthropic, resigned roughly two weeks earlier and claimed that AI companies were gambling with lives. Anthropic CEO Dario Amodei subsequently published an essay advocating for a slower pace of foundational model development and calling for third-party evaluators to audit risks such as cyberattacks and biological weapons. OpenAI CEO Sam Altman and Tesla and SpaceX CEO Elon Musk both publicly backed Amodei's proposal.

Independent AI evaluation remains in an early stage without shared criteria for deep inspections. In response, a coalition of AI evaluators is pushing frontier model developers to adopt minimum standards, including deeper system access and whistleblower protections against retaliation when unfavorable audit reports are published.