Skip to content
NewsRegulation

OpenAI proposes RSI governance framework ahead of Altman's UN Security Council briefing

· by Pondero Newsdesk

The short version

OpenAI published a formal policy proposal on September 21 for international recursive self-improvement governance standards, two days before CEO Sam Altman is scheduled to brief the UN Security Council on artificial intelligence and international security.

OpenAI proposes RSI governance framework ahead of Altman's UN Security Council briefing

OpenAI published a formal policy essay on September 21, 2026, calling for an international framework to govern recursive self-improvement in AI systems. Two days later, CEO Sam Altman is scheduled to brief the UN Security Council on artificial intelligence and international security during the UN General Assembly's high-level week in New York.

What

The essay, titled "Building standards for the next phase of AI," presents a two-part proposal. The first calls for building international standards through the emerging network of national AI safety institutes, already active in ten countries including Australia, Canada, Germany, France, Kenya, Japan, South Korea, Singapore, India, and the United Kingdom, per the September 21 essay. The coordination hub would be the Center for AI Standards and Innovation (CAISI) together with the NIST-launched International Network for Advanced AI Measurement, Evaluation, and Science. The resulting standards would establish agreed baselines for capability measurement and safeguard sufficiency. National governments would then decide whether and how to incorporate those baselines into their own legal systems. Per OpenAI, the proposal is not a mandatory pre-release review mechanism or a licensing regime.

The second part targets what OpenAI called the fragmentation problem. Without shared definitions of incidents and capability thresholds, individual national regimes produce evidence that cannot be compared across borders. OpenAI proposed a common layer of measurement and incident-reporting protocols specifically for recursive self-improvement: standards covering evaluation of RSI-relevant capability gains, thresholds for mandatory human review of automated AI research processes, and common incident severity levels for alignment failures.

Recursive self-improvement, the process by which AI systems increasingly take over the work of developing the next generation of AI, is the central concern. OpenAI acknowledged that fully autonomous RSI is not happening today, but argued that coordinated guardrails are urgent before it is. The essay also flagged the US-China angle, stating that upcoming bilateral talks on frontier AI safety "would be a positive step" and suggesting the RSI governance conversation is necessarily bilateral between the two largest AI powers.

Three senior OpenAI voices set up the September 21 essay in a coordinated sequence over the prior two weeks. Chief scientist Jakub Pachocki published an essay titled "An Alien Mind" on September 6, writing that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer," per the original essay. Chief global affairs officer Chris Lehane reversed the company's long-held position against binding US federal rules on September 9, endorsing mandatory testing, independent model assessment, cybersecurity requirements, and safety incident reporting for the most powerful models, with open-source systems exempt. The September 21 essay connected both threads to the international standards argument and named the Hugging Face incident OpenAI disclosed earlier in 2026 as "a preview of the kinds of risks that could become much more severe without robust safeguards and alignment."

OpenAI API users may face third-party assessments

The timing is deliberate and the stakes for enterprise operators are concrete. For teams running OpenAI APIs in production, the essay is the clearest signal yet that a binding international compliance layer is OpenAI's preferred direction. If national governments adopt the proposed standards, enterprise deployments of the most powerful models could eventually face third-party assessment requirements and mandatory incident disclosure obligations that sit above any individual company's internal review process. Security and legal teams at companies building on frontier models should start tracking this now, not after a UN resolution is in draft.

The proposal also puts Altman's September 23 Security Council appearance in sharper context. The session is framed as an open meeting on AI and international security. Per AI Weekly's September 21 report, Altman's remarks are expected to focus on OpenAI's safety steps and the case for shared global standards. The September 21 essay functions as the written brief he arrives with.

For competing labs, the move creates a public-positioning challenge. OpenAI explicitly called for standards that "do not advantage particular companies, countries, or business models." Anthropic and Google have not yet published equivalent RSI governance proposals. Silence after a Security Council briefing by a peer CEO is harder to sustain than silence before one.

One limitation applies throughout: all of the key claims here originate from OpenAI itself. Pachocki's assertion that no lab has solved alignment comes from the lab's own chief scientist. OpenAI has not opened its alignment research to independent external review, so these self-assessments cannot be verified by outside parties.

Context and reactions

Altman's UN Security Council briefing is the first time an OpenAI CEO has addressed that body. The session falls during UNGA high-level week, which draws heads of government and senior ministers from around the world. Past AI-related Security Council discussions have been informational rather than legislative, but an appearance there still carries significant signal weight for how member states perceive the urgency of the issue.

The essay builds on a broader OpenAI governance push throughout 2026. The company has published a proposed US federal AI framework, a misalignment reporting framework, and a research-acceleration transparency report, each cited in the September 21 essay as early contributions to the standard-setting work it now wants to internationalize. OpenAI drew an explicit analogy to aviation and financial stability, sectors where countries developed common technical standards and cooperative channels without surrendering national authority.

What to watch next

Altman's actual remarks and any formal joint statement from member states after the September 23 session are the immediate signals. A non-binding Security Council resolution on AI governance, even a procedural one, would represent a shift in multilateral framing. From the industry side, the key question is whether Anthropic or Google publish their own RSI governance proposals before or shortly after UNGA high-level week closes. If neither does, OpenAI will have established the only public benchmark on the table going into whatever standard-setting process the session produces.

Sources