The AI policy window is open. We need to act.

In a Sept. 9, 2026 essay, OpenAI Chief Global Affairs Officer Chris Lehane argues that a new chapter in AI capabilities demands a new chapter in AI policy—and that waiting for perfect bills is now the bigger risk. OpenAI says it will push Congress for mandatory, capability-based national AI safety regulation; keep backing state laws until federal action lands; work with other frontier labs on voluntary monitoring standards; and advocate internationally for compatible ways to measure capabilities, preserve human control, and decide when development should slow or stop.

The same day, OpenAI endorsed four California bills headed to Governor Newsom: SB 813 (independent safety-assessment infrastructure), AB 1405 (AI-auditor standards), SB 1119 (youth protections for companion chatbots), and AB 1864 (gene-synthesis screening against AI-enabled biological threats). Some are bills OpenAI did not previously endorse, reconsidered after a recent jump in capabilities.

Tied to Astra and Chief Scientist Jakub Pachocki's call for "extreme caution" around recursive self-improvement, the post frames a closing "defenders window" for institutions as well as for cyber defenders.

We're reading this as OpenAI trying to set the tempo of regulation before Congress goes home and before state patchworks harden into chaos. The useful core is capability-based federal rules aimed at a handful of well-resourced frontier labs—not a dragnet on startups—and an explicit refusal to launder frontier safety into a ban on open weights. That distinction matters if the goal is democratic oversight without entrenching only the largest closed labs.

I'm watching the incentives. A lab that just shipped a model it says hits Critical cyber capability has every reason to demand mandatory rules that match its preparedness story and to endorse youth and biosecurity bills that show seriousness. That does not make the endorsements fake. SB 1119's age assurance, audits, and parental controls, and AB 1864's synthesis screening, are concrete harm-reduction. Independent assessors under SB 813 and auditor registration under AB 1405 could reduce pure self-policing—if assessors get real access and are not captured.

Lehane's line on recursive self-improvement is carefully hedged: fully autonomous RSI is not happening today and should not be pursued until it can be done safely, while AI is already accelerating research tasks that used to take skilled researchers days. Governments, he argues, need shared metrics and safety bars that can force slowdowns. Our desk agrees that private voluntaryism alone is insufficient at this scale. The downside of OpenAI-shaped "industry standards" is obvious: reporting regimes designed around one lab's incident vocabulary can become the de facto federal template.

Advocate for useful AI where the essay does: medicines, infrastructure, science. Name the trajectory if the window closes unused—frontier labs keep writing their own rules while capabilities diffuse, including through open models approaching today's frontier. Bias toward action is right; surrendering the pen entirely to the labs writing the blueprint is not.

OpenAI cites prior support for California SB 53, New York's RAISE Act, Illinois SB 315, and Massachusetts audit proposals, calling state convergence "reverse federalism" until Congress codifies a baseline. Astra-era safeguards mentioned include trajectory monitoring and an alignment-evaluation gate before broader internal deployment.