House bill would bar deploying the most capable AI models until the NSA evaluates them for cyber and weapons risks

US Congress · Bill text

Read the source (opens in a new tab)
What happened
Representatives Josh Gottheimer and Mike Lawler introduced the American AI Security Act (H.R. 10981) on November 10, 2026; it was referred to Energy and Commerce and the Intelligence Committee.
Key points
  • Commerce, with the NSA Director, would define covered models by a parameter threshold and by capabilities that pose serious national-security risk or materially aid chemical, biological or radiological weapons.
  • Before deployment, developers would give the NSA full model access plus training, safety-test and deployment details; the NSA would evaluate cyber and weapons capabilities and recommend safeguards.
  • Deployment would wait until the developer adopts those recommendations or the NSA misses its deadlines (30 days to evaluate, extendable by 30, then 30 days to report); penalties are $500,000 per day, and a new court would hear developer appeals.
  • States could not impose transparency, disclosure or reporting duties on covered developers about safety protocols, safety and security frameworks, capability evaluations or risk assessments.
Evidence
Bill text as introduced. It is a proposal that has not passed; Commerce would write the rules within six months of enactment.
Uncertain
The parameter threshold and qualifying capabilities are left to those rules, so which models would be covered is not yet known.

See it in the latest record