"The AI policy window is open. We need to act." is a post published September 9, 2026 on openai.com under the company's Global Affairs category, by Chris Lehane, chief global affairs officer at OpenAI. It sets out four commitments: working with Congress on mandatory, capability-based national AI safety regulation; continuing to support state legislation until Congress acts; advancing industry-led frontier AI standards "with or without government support"; and advocating compatible international approaches to measuring capabilities and determining when development should slow or stop.
The post is a company policy statement rather than a research document. Its empirical premises — the capabilities of Astra, and the claim that AI is already accelerating parts of the research used to develop and align the next generation of models — rest on OpenAI's own earlier publications rather than on evidence presented with the post.
Framing
Lehane writes that the company has "reached a new chapter in AI capabilities, and that demands a new chapter for AI policy," that "no company, industry, or government can meet this challenge alone," and that the moment should be met "with a bias toward meaningful action over policy perfection." The post opens by citing chief scientist Jakub Pachocki's essay "An Alien Mind", published three days earlier, which called for "extreme caution" in light of the potential for recursive self-improvement, and states that technical work inside individual laboratories "will not be enough."
The post borrows a framing from OpenAI president Greg Brockman's "defenders window" — described as a limited period in which frontier AI can help defenders strengthen critical systems before offensive capabilities become widespread — and asserts an analogous closing window for policymakers "to establish durable safeguards before AI capabilities outpace the institutions responsible for governing them." Lehane writes that "as capabilities grow, confidence in safety must increasingly set the pace of AI progress."
The post also states that capabilities "will not remain confined to a few frontier laboratories," and that models developed around the world, including open models, "will increasingly approach today's frontier and become broadly available."
Disclosed safeguards
The post describes safeguards OpenAI says it has added across the model-development lifecycle: stronger isolation for frontier research workloads, expanded monitoring of model behavior during tool-enabled training and evaluations, and clearer rules for when to escalate concerns. For Astra specifically, it states that OpenAI introduced universal monitoring of full trajectories, including chains of thought, and a mandatory alignment-evaluation gate before broader internal deployment. Lehane restates the commitment in OpenAI's Preparedness Framework to slow or stop development or deployment of systems the company cannot sufficiently safeguard, "as we have done before."
The trajectory-monitoring disclosure sits against Pachocki's statement in "An Alien Mind" that OpenAI's ability to rely on chain-of-thought monitoring is progressively diminishing. See Reasoning Models and Chain-of-Thought.
Recursive self-improvement
The post states that fully autonomous recursive self-improvement — "in which AI systems independently drive successive generations of increasingly capable AI" — "is not happening today," and that "we should not pursue it unless and until it can be done safely." It adds that AI is already accelerating parts of the research used to develop and align the next generation of models, citing OpenAI's research-acceleration disclosure finding that agents can perform some tasks that would take skilled researchers several days, and characterizes this as "evidence of the direction of travel" rather than recursive self-improvement itself.
Lehane states OpenAI's aim as safely building automated AI researchers that work under human supervision "to advance both deep learning and alignment," and asks governments to develop common ways to measure progress toward recursive self-improvement, preserve meaningful human control, and set shared safety bars for when development should slow or stop. Where safety bars cannot be met without slowing capability growth, the post says the safety bars should be prioritized. See Recursive Self-Improvement (RSI).
Federal position
The post states that "the prospect of AI-accelerated AI development demands more than voluntary commitments" and that the United States "needs mandatory, capability-based national regulation that can evolve as the technology does." It refers to OpenAI's Blueprint for Democratic Governance of Frontier AI for the framework's elements: common testing and independent-assessment requirements, stronger cybersecurity protections, incident-reporting rules, national preparedness, and shared measures for tracking progress toward recursive self-improvement. Lehane writes that "several serious frontier safety proposals are now taking shape in Congress," that OpenAI expects to support legislation that "materially raises the safety bar," and that "Congress should act before it adjourns."
Four constraints are placed on the framework the post asks for:
| Constraint | As stated |
|---|---|
| Scope | Frontier safety requirements should apply to "the handful of well-resourced laboratories developing the most capable systems," not startups, small developers or researchers "operating nowhere near the frontier"; obligations proportionate to capabilities and risks |
| Open weights | "Nor should frontier safety policy become open-weights policy by another name"; open models can be part of the solution, particularly in cybersecurity and where sovereignty, security or data-residency needs favor local deployment, and most compete "on cost, control, and latency" rather than at the frontier |
| Competition | A federal framework should address frontier capabilities and risks "without weakening competition, entrenching incumbents, or driving innovation overseas" |
| Concentration of power | A public framework "should reduce, not increase, the concentration of power"; democratically accountable standards, independent verification and meaningful transparency "would replace" the present arrangement in which frontier laboratories "largely set their own rules" |
The post cites OpenAI's signature on the Open Weights and American AI Leadership letter, hosted by NVIDIA, for the proposition that "America needs both open and closed models." See Open-Source AI / Open-Weight Models.
Lehane closes the federal argument by asserting that "the United States cannot win the AI era on capability alone" and that being technically ahead "will matter little if people, institutions, and governments do not trust the technology enough to adopt it."
State endorsements
The post restates OpenAI's support for California SB 53, New York's RAISE Act, Illinois SB 315, and independent audits in the frontier-safety legislation under consideration in Massachusetts, and encourages states to converge on those safeguards so as to "create a de facto national baseline that Congress can ultimately codify — an approach we call reverse federalism," linking to the earlier OpenAI post that introduced the term. It adds that "harmonization does not mean freezing requirements in place." The term is OpenAI's own; see Reverse Federalism.
It then announces formal endorsement of four additional California bills that had passed the legislature and were headed to Governor Gavin Newsom:
| Bill | As described in the post |
|---|---|
| SB 813 | Would establish a process for designating qualified, independent organizations capable of assessing AI risks |
| AB 1405 | Would create registration, independence, transparency and accountability requirements for AI auditors |
| SB 1119 | Would require age assurance, risk assessments, independent audits, parental controls and safeguards against harmful content for children and teens using companion chatbots |
| AB 1864 | Would require gene-synthesis providers and manufacturers of benchtop synthesis equipment to follow federal screening standards |
On SB 813 the post states a preference for independent technical assessments to be required at the federal level, "where policymakers can establish consistent standards for assessor qualifications, security, and access to highly sensitive information," and supports the state measure only "in the absence of federal action." On SB 1119 it notes OpenAI had already signaled support, and situates the endorsement alongside the company's youth-safety product measures — stronger defaults in sensitive areas, Quiet Hours, Study Hours and break reminders in ChatGPT for Teens — and its support for the Parents & Kids Safe AI Act.
The post records a change of position: "Some of these bills we did not endorse in the past, and are now supporting after reconsidering in light of the recent jump in capabilities we have seen." It does not identify which of the four were previously opposed or unendorsed. It adds that the bills "are not a substitute for federal regulation."
Industry-led monitoring standards
Lehane argues that "laws alone will not make advanced AI safe" and names monitoring as "the most urgent place to start," particularly for misalignment, defined in the post as a model pursuing "an objective in ways that violate human intent or established boundaries, without requiring consciousness or malicious intent." As models gain autonomy, use more powerful tools and operate over longer periods, the post says developers "should be required to monitor for these misaligned behaviors and demonstrate that effective safeguards are in place."
Monitoring is tied to a disclosure rule the company says it advocated in California: that companies "should be required to provide prompt written notice to affected parties when, during development or evaluation, their models circumvent another organization's security controls without authorization and materially access, alter, or destroy that organization's protected systems or confidential information." OpenAI also states support for federal reporting requirements for other serious AI incidents, and says it is working to define which incidents should be covered. The described rule tracks the circumstances of the Hugging Face evaluation-security incident disclosed by OpenAI earlier in the same period; see AI Incident Reporting.
The post refers to a framework OpenAI said it was developing "as we shared last week" for reporting consequential misalignment incidents and systematically monitoring frontier-model activity, including internal use, beginning as a company-led effort intended to "help inform broader federal policy and reporting requirements." It states that OpenAI wants to work with other laboratories on industry-led standards, that such standards "would complement — not replace — mandatory federal safeguards and democratic oversight," and that "everything should be on the table — including considering what can be built voluntarily outside government as an immediate step."
OpenAI published that framework on September 16, 2026 (Our Framework for Reporting Model Misalignment (OpenAI, September 2026)).
On the international dimension the post argues that standards "must ultimately also extend beyond national borders," since models, research and expertise move globally and failures involving the most capable systems "could have consequences far beyond the country in which they are developed," and states that OpenAI is "increasingly convinced that compatible international standards will be necessary." See AI Governance (umbrella).
Documents referenced
The post links to several documents not held here: Greg Brockman's "defenders window" post, OpenAI's Preparedness Framework v2 PDF, "A Blueprint for a Federal Framework," and the Open Weights and American AI Leadership letter hosted by NVIDIA. It also links to OpenAI's introductions of GPT-Rosalind, Daybreak for Frontline Defenders, and its economic-opportunity and national-science posts as illustrations of claimed benefits.
Relationships
- related: Building Standards for the Next Phase of AI (OpenAI, September 2026) — restates this post's RSI position and proposes international standards through CAISI and the AI safety institutes.
- supports: An Alien Mind (Jakub Pachocki, OpenAI, September 2026) — the post cites the chief scientist's essay as the premise for its policy urgency, and extends its call for shared safety bars into specific legislative asks.
- depends-on: Democratic Governance of Frontier AI: A blueprint for a federal framework (OpenAI, June 2026) — the federal framework the post advocates is the blueprint's content.
- related: Keeping America out in front on AI (OpenAI Global Affairs, August 2026) — earlier statement of the same state-convergence-to-federal-codification argument.
- related: Reverse Federalism — the term the post uses for its state strategy, attributed to OpenAI.
- related: California SB 813 (AI independent verification organizations), California AB 1405 (AI auditor registry), California SB 1119 (companion chatbot safeguards for minors), California AB 1864 (gene-synthesis and benchtop-synthesis screening) — the four endorsements announced.
- related: Astra — the safeguards disclosed are Astra-era.
- related: Recursive Self-Improvement (RSI) — the post's stated position on pursuit and measurement.
- contradicts: David Sacks — the post asks Congress for mandatory federal requirements, against the administration AI adviser's position that federal action should preempt rather than impose frontier obligations.
- related: Our Framework for Reporting Model Misalignment (OpenAI, September 2026) — the monitoring and disclosure framework the post announced