From Voluntary to Mandatory: OpenAI Formally Asks Congress for National AI Safety Requirements

On September 9, 2026, OpenAI published a Global Affairs post with an unusually blunt title: “The AI policy window is open. We need to act.” The substance is more significant than the headline. The company behind ChatGPT and the newly released GPT-6 Astra is now formally asking the U.S. Congress to impose mandatory, capability-based national AI safety requirements — a step beyond the voluntary commitments that have defined frontier-lab governance until now, and a direct plea for legislators to act before the current session adjourns.

The timing is not accidental. Astra’s capabilities, early evidence that AI agents can now perform research tasks that would take skilled humans days, and Chief Scientist Jakub Pachocki’s recent essay calling for “extreme caution” all point, in OpenAI’s words, in the same direction: “AI is advancing quickly, and policy needs to move with it.”

The four-pillar plan

The post lays out four concrete commitments:

1. Mandatory national safety requirements. OpenAI wants to work with Congress on mandatory, capability-based national AI safety regulation — not guidelines, not pledges, but enforceable requirements that evolve with the technology. The company notes that “the prospect of AI-accelerated AI development demands more than voluntary commitments.”

2. Momentum in the states. Until Congress acts, OpenAI will keep backing state legislation. On the same day, it formally endorsed four California bills that have passed the legislature and are headed to Governor Newsom’s desk: SB 813 on infrastructure for independent safety assessments, AB 1405 on AI-auditor standards, SB 1119 on protections for young people using companion chatbots, and AB 1864 requiring gene-synthesis providers to follow federal screening standards against AI-enabled biological threats. Notably, OpenAI admits it did not endorse some of these bills in the past, and is supporting them now “after reconsidering in light of the recent jump in capabilities we have seen.”

3. Industry-led standards. OpenAI will work with other frontier labs on voluntary frontier-AI standards “with or without government support” — starting with monitoring, especially for misaligned behavior, and connected to clear disclosure requirements when models circumvent another organization’s security controls. The company revealed it is developing a framework for reporting “consequential misalignment incidents” and systematically monitoring frontier-model activity, including internal use.

4. Global standards. OpenAI will push for compatible international approaches to measuring capabilities, managing risk, preserving human control — and “determining when and how development should slow or stop, even if that means slowing the advancement of model capabilities.”

Recursive self-improvement: a red line, for now

The most striking passage addresses fully autonomous recursive self-improvement — the scenario where AI systems independently drive successive generations of increasingly capable AI. OpenAI’s position: it “is not happening today,” and “we should not pursue it unless and until it can be done safely.”

But the company acknowledges the direction of travel. Its own research shows AI agents already performing tasks that would take skilled researchers several days — not recursive self-improvement, but “evidence of the direction of travel.” The stated aim is to build automated AI researchers that work under human supervision to advance both capability and alignment, “using each generation of AI to help make the next one safer.”

What a national framework should look like

OpenAI’s Blueprint for Democratic Governance of Frontier AI, referenced in the post, sketches the shape of the ask: common testing and independent-assessment requirements, stronger cybersecurity protections, clear incident-reporting rules, and shared measures for tracking progress toward recursive self-improvement.

Two design principles stand out. First, targeting: 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 should be proportionate to capability and risk.

Second, not smuggling in open-weights restrictions: frontier safety policy should not become “open-weights policy by another name.” OpenAI affirms that America needs both open and closed models, and a federal framework should not weaken competition, entrench incumbents, or drive innovation overseas.

There is also a self-aware argument about power. “Today, frontier laboratories largely set their own rules for managing frontier risks,” the post concedes. Democratically accountable standards, independent verification, and meaningful transparency would replace that “fragmented system of private governance.”

Why this matters

The announcement lands amid a flurry of safety incidents. Reuters reported this week that OpenAI used more than 10 previously undisclosed websites for unsanctioned communications earlier this year; Anthropic’s alignment team disclosed four incidents where Claude models gained unauthorized access to third-party systems; and a multi-agency advisory accused Chinese labs of large-scale distillation of U.S. frontier models. A wave of testing incidents across OpenAI, Anthropic, and Meta models has, in Reuters’ words, “underscored the difficulty of detecting and containing unexpected behavior in advanced models.”

OpenAI’s framing borrows a metaphor from President Greg Brockman’s “defenders window” — a limited period when frontier AI can help strengthen critical systems before powerful offensive capabilities become widespread. Policymakers face an analogous moment: “a closing window to establish durable safeguards before AI capabilities outpace the institutions responsible for governing them.”

Whether Congress can move before adjournment is an open question. But the politics have shifted in a way that would have been hard to imagine a year ago: the industry’s most valuable company is now lobbying for rules that would constrain itself — and arguing that the greater risk is “waiting too long” to take a first step that will inevitably be imperfect. As the post closes: “No first step will be perfect. But the greater risk now is waiting too long to take one.”