OpenAI has published a detailed proposal for international technical standards to govern frontier AI development, with a particular focus on the risks of recursive self-improvement (RSI). The September 21, 2026 statement outlines how automated AI research - where systems take on more of the work of building next-generation AI - could accelerate progress beyond collective human oversight without shared safety baselines and incident reporting protocols across countries.
The proposal arrives as AI-enabled research has already led to advances in mathematics, including the Navier-Stokes Millennium Problem, and as multiple nations operate their own AI safety institutes. OpenAI frames the moment as a decision point: the United States can lead a coordinated standards effort or watch a fragmented system take hold.
The RSI challenge and why it changes the equation
Automated AI research can involve varying degrees of human supervision. As systems take on more of the work of developing successive generations of AI, they can increasingly drive a process of recursive self-improvement, even while people remain involved. As this process becomes more automated, the pace of AI progress could accelerate rapidly.
OpenAI is direct about the stakes. "Done without appropriate care and caution, RSI could result in humans losing practical control over AI development, unable to provide oversight on research processes they no longer understand," the statement reads. The company points to the previously disclosed Hugging Face Incident as "a preview of the kinds of risks that could become much more severe without robust safeguards and alignment." Fully autonomous RSI is not happening today, and OpenAI said it "should not pursue it unless and until it can be done safely."
The company also argues that automated AI research, done right, could help solve alignment problems. An automated AI researcher can also be an automated AI safety researcher, capable of strengthening defenses, securing critical infrastructure, and developing new protective measures.
Three problems that international standards would address
OpenAI identifies three structural challenges that make national-level action alone insufficient. First, fragmentation - evaluations, reporting requirements, and incident definitions that differ by nation make it harder to compare evidence and respond to cross-border risks. Second, collective action - each nation acting independently can produce outcomes no nation wants, especially if RSI accelerates research beyond the ability to assess risks. Third, uneven capacity - frontier development expertise is concentrated in a small number of countries, which compounds the other two problems.
These challenges apply to both open and closed models, the company said. The rationale for standards is "rooted in avoiding the concentration of power, and producing better practical outcomes." Standards provide a way for stakeholders outside of AI labs to have a say in how the technology unfolds.
A two-part mechanism for global coordination
The proposal outlines two essential components. The first is a mechanism that facilitates complementary national and international frontier standards, using the emerging network of AI safety institutes already established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, and the United Kingdom. These institutes could facilitate standard-setting through the Center for AI Standards and Innovation (CAISI) and national industry bodies.
Standards developed through this effort would provide a common technical foundation for capability measurement, risk assessment, and safeguard sufficiency. They would not be licenses or mandatory prerelease review requirements. National governments would decide whether and how to incorporate them into their own legal systems. OpenAI said the work should be developed transparently and designed so it does not advantage particular companies, countries, or business models, "including by making it harder for new entrants or open-weight developers to compete."
The second component focuses on common measurements and incident reporting protocols. These would include standards for evaluating RSI-relevant AI progress, defining what kinds of automated AI research processes should trigger immediate human review, and establishing incident classification, tracking, and reporting thresholds. OpenAI's recent report on research acceleration and its misalignment reporting framework are offered as early contributions to this effort.
The proposal also calls for secure channels of communication between critical infrastructure operators and governments worldwide to share national security concerns, emerging vulnerabilities, and best practices. Dialogue between the United States and China in these areas "would be a positive step," the statement said.
Why this matters for executives, policy professionals, and legal teams
For leaders in strategy, government, and legal roles, the proposal signals that the standards conversation is shifting from voluntary principles toward technical specifications with real operational teeth. The emphasis on incident classification and reporting thresholds - not just capability benchmarks - suggests future compliance frameworks will require demonstrable evidence of safety practices, not stated commitments. Professionals responsible for AI governance, risk assessment, or public policy should watch whether CAISI and the international network of safety institutes begin producing draft standards, as those technical documents will shape what regulators and business partners eventually expect. Those building internal expertise can look to structured learning resources like AI Public Policy Courses and AI Safety Engineering Courses to prepare their teams for the technical and regulatory demands ahead.
Your membership also unlocks:
