OpenAI calls for U.S.-led global standards on frontier AI

On Sept. 21, 2026, OpenAI published “Building standards for the next phase of AI,” calling for a U.S.-led international effort to set global technical standards for frontier AI — including recursive self-improvement (RSI) — through networks of AI safety institutes and common measurement and incident-reporting protocols.

OpenAI says automated AI research is speeding its own work and that fully autonomous RSI is not happening today and should not be pursued until it can be done safely. The ask is shared standards so labs and countries can define “what good looks like” on catastrophic-risk mitigation — without OpenAI treating those standards as licenses or prerelease approvals.

The rule in English

RSI frame. Automated AI research can involve varying human supervision; as systems do more of successive-generation development, they can drive RSI even while people stay involved. Done carelessly, OpenAI says, RSI could leave humans without practical control. Hugging Face incident cited as a preview of risks that get worse without safeguards (not claimed as RSI itself).

Why standards. International technical standards for safety/security practices may matter as much as alignment research for pacing the frontier. Challenges named: fragmentation of evaluations/reporting across nations; collective-action problems as RSI accelerates research; uneven global capacity. Applies to open and closed models.

Ask (1) — national/international mechanism. U.S. should lead with other countries. One path: leverage AI safety institutes (examples listed: Australia, Canada, Germany, France, Kenya, Japan, Korea, Singapore, India, UK) via CAISI and national industry bodies, focused on (a) frontier models/developers by capability benchmarks and (b) benefit-risk management for automated AI research, including RSI. Build on CAISI’s International Network for Advanced AI Measurement, Evaluation, and Science (2024). Standards = common technical foundation for capability measurement, risk assessment, safeguard sufficiency — not licenses, mandatory prerelease review, or approval requirements. Governments decide whether/how to write them into law. Consult open- and closed-model developers; avoid entrenching incumbents.

Ask (2) — measurements and incident protocols. Standards for: evaluating RSI-relevant progress and autonomous research share inside labs; human oversight triggers for automated research; incident classification/tracking/reporting/response for alignment and automated-research issues (severity levels, thresholds). OpenAI points to its research-acceleration reporting and misalignment reporting framework as early contributions. Also: secure channels for critical-infrastructure operators and governments; U.S.–China dialogue flagged as positive if it happens.

Pacing definition. Not a fixed speed — alignment/deployment of safeguards staying ahead of capabilities; standards and institutions amplifying market and societal forces.

Who wrote it / who it binds

Writer: OpenAI. Binds no government and no rival lab. A nonbinding policy request aimed at Washington and institute networks.

Who it exempts

Explicitly: these technical standards would not themselves be licensing or prerelease approval regimes. Open-weight entrants should not be disadvantaged by design, OpenAI says.

What a founder, parent, or worker should do differently

Frontier founders: treat RSI measurement, autonomous-research share, and incident thresholds as the next disclosure bar OpenAI wants written internationally — prepare to report them.

Policymakers: OpenAI wants U.S. leadership through CAISI/institutes, not a UN-only path — and still rejects treating the standards as permits.

Parents and workers: not a school or app rule — geopolitical/lab governance speech.

What’s next

• Whether CAISI or peer institutes pick up RSI measurement / incident protocols.
• Bessent–China AI talks vs. this ask for secure channels.
• Contrast with Trump rejection of new federal guardrails (prior coverage) and Amodei/Anthropic pacing essay.
• “OpenAI again cites math progress including Navier–Stokes, covered in our Sep 21 5pm math-advisory story.”

Source: OpenAI, “Building standards for the next phase of AI,” Sept. 21, 2026.