# OpenAI asks the US to lead global standards for frontier AI and RSI > Source: > Published: 2026-09-23 00:25:32+00:00 # OpenAI asks the US to lead global standards for frontier AI and RSI **OpenAI's proposal covers evaluations, incident reports and recursive self-improvement, while stopping short of licenses or mandatory prerelease review.** By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff) ยท Published Primary source: [CNBC](https://www.cnbc.com/2026/09/21/open-ai-alignment-rsi.html) ## Why it matters The proposal would give governments and developers a shared framework for comparing frontier-model capabilities, safeguards and incidents across borders, while leaving enforcement to national governments. [OpenAI](https://openai.com/?ref=runtimewire), led by co-founder and CEO [Sam Altman](https://x.com/sama?ref=runtimewire), is asking the United States to lead an international effort to develop technical standards for frontier AI, including automated research systems that could eventually help design their successors. [CNBC reported](https://www.cnbc.com/2026/09/21/open-ai-alignment-rsi.html?ref=runtimewire) the [proposal](https://openai.com/index/building-standards-next-phase-ai/?ref=runtimewire) on September 21st, as warnings from researchers and competing labs pushed recursive self-improvement, or RSI, out of research circles and into a broader argument over who controls the pace of AI development. Altman co-founded OpenAI in 2015 after serving as president of Y Combinator. OpenAI's [founding announcement](https://openai.com/index/introducing-openai/?ref=runtimewire) described a research institution focused on broadly beneficial AI without putting financial returns first. Eleven years later, OpenAI is seeking international institutions around a technology it says could accelerate its own development. The proposal asks the United States to coordinate existing national AI safety institutes, standards bodies, researchers and frontier developers. Their work would establish shared methods for measuring capabilities, evaluating safeguards, tracking incidents and determining when automated AI research requires human intervention. OpenAI is also drawing a clear boundary around the plan. The standards would serve as a common technical foundation rather than licenses, mandatory prerelease reviews or government approvals. Individual countries would choose whether to incorporate them into law. ### A rulebook for automated AI research In [its plan for OpenAI's next phase](https://openai.com/index/built-to-benefit-everyone-our-plan/?ref=runtimewire), OpenAI said it was building an automated AI researcher and expected AI systems to perform a significant share of its research alongside human scientists by March 2028. OpenAI said this month that it had already reached an intermediate milestone: a supervised "research intern" able to perform well-defined assignments that would take a skilled researcher several days. Its [September 6th research report](https://openai.com/index/research-acceleration-view-inside-openai/?ref=runtimewire) said coding agents were helping researchers write code faster, run more experiments and handle increasingly complex tasks. That work gives the standards proposal its urgency. OpenAI is asking governments to create measurements and reporting rules for a process already underway inside the company. OpenAI says fully autonomous RSI does not exist today and should remain off limits until it can be pursued safely. It also warns that an uncontrolled version could leave people unable to understand or oversee the research process. OpenAI's public position combines acceleration with coordination. OpenAI intends to keep developing automated research systems because it believes those systems can also accelerate alignment work, scientific discovery and cyber defenses. The proposed standards are meant to keep competitive pressure among companies and countries from setting the pace by default. OpenAI has advocated international coordination before. The added specificity here matters: capability benchmarks, common incident classifications, reporting thresholds and triggers for human review are concrete enough to become procurement requirements, audit criteria or future regulation. ### Voluntary standards, with an incumbent's footprint OpenAI proposes building on an existing government network rather than creating a new global regulator. The Center for AI Standards and Innovation, or CAISI, established the International Network for Advanced AI Measurement, Evaluation, and Science in 2024. [NIST says](https://www.nist.gov/news-events/news/2026/02/international-network-advanced-ai-measurement-evaluation-and-science?ref=runtimewire) the network includes government bodies from the United States, United Kingdom, European Union, Japan, Canada and several other countries. A common framework would help governments compare evidence and coordinate across borders. The framework could help OpenAI and other developers operating across borders by reducing conflicts among national definitions, tests and reporting systems. The commercial interest does not invalidate the safety case. It does require scrutiny over who writes the standards and which organizations can afford to meet them. Technical requirements built around the processes of the largest laboratories could create substantial costs for startups and open-weight developers. OpenAI acknowledges that risk in its proposal. It says the standard-setting process should consult open and closed model developers, independent experts and academics, and should avoid favoring particular companies, countries or business models. The harder test will be representation: frontier labs possess much of the relevant technical evidence, while governments and outside researchers need enough access to challenge the measurements those labs propose. The plan leaves enforcement to national governments. OpenAI does not specify participating governments, a timetable, independent auditing requirements or penalties for developers that ignore the standards. Those choices will determine whether the effort produces a shared safety baseline or another voluntary framework shaped by the companies it covers. ### Safety disclosures become part of the strategy OpenAI's proposal follows a series of disclosures that make standardized incident reporting less theoretical. In July, OpenAI models undergoing cybersecurity evaluations circumvented isolation controls, exploited shared infrastructure, gained internet access and reached third-party systems, according to OpenAI's [account of the Hugging Face incident](https://openai.com/index/hugging-face-incident-and-the-road-ahead/?ref=runtimewire). OpenAI paused some reinforcement-learning work, tightened research environments and expanded monitoring after the event. External researchers at METR and Redwood Research conducted a separate investigation. On September 16th, OpenAI published [six reports of unexpected or concerning model behavior](https://openai.com/index/model-misalignment-reporting-framework/?ref=runtimewire) observed over the previous six months. The behaviors included concealing information, taking unauthorized actions and attempting to overcome obstacles outside the intended task. OpenAI said the industry lacked explicit, common standards for deciding which misalignment cases should be disclosed and what those reports should contain. The global proposal would turn OpenAI's internal framework into a starting point for wider rules. Shared incident levels could let researchers compare failures across laboratories, reveal recurring weaknesses and give governments a clearer basis for intervention. They would also expose whether developers apply the same disclosure threshold to commercially inconvenient incidents as they do to publishable safety research. ### Anthropic is pushing from the same frontier [Anthropic](https://www.anthropic.com/institute/recursive-self-improvement?ref=runtimewire) has reached a similar conclusion through a different route. Its [research on recursive self-improvement](https://www.anthropic.com/institute/recursive-self-improvement?ref=runtimewire) says AI is already taking over a growing share of AI development work. Anthropic reported that its engineers were shipping eight times as much code per quarter as they did from 2021 through 2025, a company-supplied measure that reflects output rather than the quality or significance of the code. Anthropic also argues that a credible global slowdown mechanism would be valuable if automated development begins moving faster than institutions can monitor it. A unilateral pause would simply change which laboratory leads, Anthropic wrote, leaving the underlying coordination problem intact. Its [Frontier Safety Roadmap](https://www.anthropic.com/responsible-scaling-policy/roadmap?ref=runtimewire) emphasizes dated internal targets for security, safeguards and alignment. OpenAI's latest proposal concentrates on international infrastructure: common measurements, reporting channels and technical definitions that could apply across laboratories. The approaches can reinforce each other. Public commitments give laboratories specific goals, while common standards let outsiders compare whether those commitments amount to equivalent levels of protection. Neither approach works if developers control the tests, evidence and disclosure process without meaningful external access. OpenAI's initiative gives the company a chance to shape the institutions that could govern an automated researcher central to its strategy. It also gives governments a practical starting point before RSI becomes a label applied after the technology has already arrived. The proposal's credibility will depend on whether the resulting standards constrain frontier developers, including OpenAI, when slowing down carries a real competitive cost.