The AI Safety Menu

All laws

Fund defensive tech

Public funding for defensive technology. Public money for early pathogen warning, pandemic protection and AI tools that patch critical systems.

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

The government funds technology that blunts AI-enabled catastrophes: a nationwide network that spots new pathogens early, better masks and air-cleaning for pandemics, and AI tools that find and patch security holes in water, power and hospital systems. This is the 'defensive acceleration' idea.

Applies to: Federal biodefense and cyber-defense programs.

Where things stand

Defensive spending exists but is small and mostly falling: CDC wastewater and traveler testing, the pandemic-medicine agency BARDA, the national stockpile and CISA. There is no large program aimed at AI-enabled threats.

Why it costs almost no lead

It spends public money, not lab time. Frontier labs lose at most a day to competition for talent.

Biggest unknown: How much defenses can keep up with AI-enabled attacks, and how quickly a new program could be built: a national pathogen-detection network would take two to three years.

Why it lowers p(doom) by ~0.15%

Makes it harder for an AI-enabled bioweapon or cyberattack to become a catastrophe, by catching outbreaks early and patching critical systems.

If AI makes attacks easier, stronger defenses cap the damage. It does nothing about AI slipping out of human control, and defense often lags offense.

The strongest case that it costs more

For AI-enabled bio and cyber threats, attack likely outpaces defense, and defensive technology does nothing about AI that is misaligned. Big spending could buy false security and pull money and attention away from preventing dangerous capabilities in the first place.

The debate

For

Against

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.