The AI Safety Menu

All laws

Warner NSA bill

Secure AI Development Act (Warner). Give the NSA's AI Security Center access to frontier models 21 days before release.

Lead costAlmost noneof America's lead over China, over 3 years
Cuts p(doom) by~0.15%from a 5% starting estimate
Holds up○ Training○ Lab's own use● Public release
Enact it on the menuSee the findings

What it does

A proposed federal law that would require labs to give the NSA's AI Security Center access to new models 21 days before release.

Applies to: Frontier models, defined by capability rather than a compute threshold.

What's in it: 4-week test period

Where things stand

Labs already share pre-release checkpoints with CAISI and UK AISI, for periods from three days to three weeks. A 21-day floor is at the long end of current practice.

Why it costs almost no lead

Only the public release waits, so the lab's own progress barely moves.

Biggest unknown: Whether 21 days of access is a floor on the release date or just a checkpoint-sharing rule.

Why it lowers p(doom) by ~0.15%

Gives the NSA three weeks to test frontier models for security dangers before release.

The strongest case that it costs more

Handing weights to the NSA three weeks before every launch is a standing intelligence relationship, not an evaluation. If the agency finds something it does not like, the release conversation happens with the government. The bill's text does not gate release on the agency's approval, but a lab would be brave to ship over its objection.

Sources

Rough starting points, not precise forecasts. Lead costs assume China doesn't depend on U.S. models, the case least favorable to safety laws, and count 3 years. On the menu you can change every assumption and put in your own numbers. Last priced 2026-09-26.