OpenAI Publishes Proposal for International Frontier AI Standards

OpenAI announced on September 21 that the United States should take the lead in building international technical standards for the safety and security of frontier AI. In the announcement the company wrote that "the United States should lead an effort to work together with countries around the world to develop global technical standards for frontier AI, including for RSI." The proposal has two pillars: a mechanism that connects national and international standards, and common measurement and incident reporting protocols.

The reason standards came up is recursive self-improvement. As AI systems take on more of the work of developing successive generations of AI, the process can increasingly drive itself even while people remain involved, and the more automated it becomes the faster progress could accelerate, the company explained. The company wrote that fully autonomous recursive self-improvement is "not happening today, and we should not pursue it unless and until it can be done safely." A warning sits in the same passage: done without appropriate care, it could leave humans without practical control over AI development, unable to oversee research processes they no longer understand.

The company wrote that the Hugging Face Incident it disclosed, while not a direct result of recursive self-improvement, is a preview of the kinds of risks that could become much more severe without robust safeguards and alignment. The announcement opens with three goals recently set out by Chief Executive Officer Sam Altman and Jakub Pachocki: navigating the next period of AI progress by building an automated AI researcher, iterating with it on the alignment problem and finding ways for people to remain part of the self-improvement loop; delivering the benefits of scientific progress and economic growth; and empowering everyone individually with a personal AGI. The company said AI is accelerating its own research and has led to advances in mathematics including the Navier-Stokes Millennium Problem.

The company named three problems that make international standards necessary. Fragmentation, where evaluations, reporting requirements and incident definitions conflict across nations, making it harder to compare evidence and to respond to risks that cross borders. Collective action, where each nation acting independently produces outcomes no nation wants. And uneven capacity, because frontier development and the expertise around it are not evenly distributed across the world. The announcement said these problems apply to both open and closed models.

The first mechanism is a proposal to use what already exists. The company wrote that the network of AI safety institutes established in Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India and the United Kingdom could be leveraged to set standards through the Center for AI Standards and Innovation and national industry bodies. The focus is twofold: frontier AI models and developers as measured by capability benchmarks, and benefit-risk management for automated AI research including recursive self-improvement. The International Network for Advanced AI Measurement, Evaluation, and Science, created by the Center for AI Standards and Innovation in 2024, was named as the starting point.

What the standards are not is nailed down in the announcement as well. The technical standards developed here would not be licenses, not mandatory prerelease review, and not approval requirements for AI models. National governments would decide whether and how to incorporate them into their own legal systems. The company wrote that common standards should not be designed to advantage particular companies, countries or business models, and should not make it harder for new entrants or open-weight developers to compete.

The precedents the company cited are aviation and financial stability, fields where countries have developed common technical standards and trusted channels for cooperation without giving up national authority. The announcement said the work should also proceed closely with established and newer standards bodies such as ISO, the Frontier Model Forum, the Agentic AI Foundation and the Open Secure AI Alliance, as well as implementation-focused organizations like the Appia Foundation, which is developing practical specifications that connect international standards with real-world AI assessments.

The second pillar is a proposal to align how things are measured and reported. The items the company listed are standards for evaluating recursive self-improvement-relevant AI progress and the amount of autonomous research happening within a company; standards for human oversight, including what kinds of automated research processes should trigger immediate human review; and standards for incident classification, tracking, reporting and response, including common severity levels and reporting thresholds. The company offered its own report on research acceleration and its misalignment reporting framework as early contributions. METAL has reported that OpenAI released a model misalignment reporting framework.

The announcement also asked for machinery outside the standards themselves. Critical infrastructure operators and governments worldwide should establish secure channels of communication to share national security concerns, emerging vulnerabilities and threats, and best practices. The company wrote that dialogue between the United States and China in these areas would be a positive step and that upcoming talks are happening at an opportune time. According to reports, Chief Executive Officer Sam Altman is scheduled to attend a United Nations Security Council meeting in person on September 23 to brief on AI and international security.

The domestic blueprint this proposal rests on is a separate document. The paper on democratic governance of frontier AI, which METAL examined, is a nine-page PDF linked from the announcement and dated June 2. It calls the approach of letting states refine a common framework first and having the federal government build on it reverse federalism, and says California's SB 53, New York's RAISE Act and Illinois's SB 315 have already converged on the core elements. The minimum requirements it lists for a federal framework are severe risk evaluations and mitigations, transparency reports, independent third-party audits, critical safety incident reporting, model weight security, whistleblower protections and meaningful accountability mechanisms.

The document calls for building the Center for AI Standards and Innovation into the world's premier institution for frontier AI evaluation, standards development, independent assessment certification and coordination across government and with international partners. Its director should report directly to the Secretary of Commerce or another senior Cabinet-level official, and it should have flexible hiring authorities similar to those used by CHIPS for America along with access to classified computing environments. The strongest demand comes last: once a comprehensive federal framework is in place, state laws that seek to regulate the same frontier safety risks should be preempted into a single national framework.

Read through a lawyer's lens, this announcement pushes in two directions at once. Internationally it builds technical standards with no approval authority and leaves each country to write them into law on its own. Domestically it asks for state laws to be preempted in exchange for a federal framework. The sentences that evaluations happen but deployment decisions stay with the developer, and that developers may deploy without penalty if an evaluation is not completed within the defined period, belong to the same design. As the company itself wrote, the rationale for standards is rooted in avoiding the concentration of power and producing better practical outcomes, and who writes those standards first will set the rules of the coming decade.