House bill would have the NSA test frontier models before release
H.R. 10981 would bar release of covered models until the NSA has evaluated cyber, chemical, biological and radiological capabilities and the developer has implemented its recommendations; it was referred to two committees.
- Government & policy
- Safety & alignment
- Security & misuse
- Minor
Representatives Josh Gottheimer and Mike Lawler introduced the American AI Security Act, H.R. 10981, on 10 November. It was referred to the House Energy and Commerce Committee and the Permanent Select Committee on Intelligence. It follows the same pair’s Stop Rogue AI Act from September.
The bill would direct the Secretary of Commerce, with the National Security Agency’s director, to issue regulations within six months of enactment defining a “covered foundation model”. That definition would include a minimum parameter threshold and any demonstrated capability that could pose a serious national-security risk, including through cyber threats, or materially help someone develop a chemical, biological or radiological weapon.
Developers of covered models would have to give the NSA full access to the model, plus information on training, safety tests and results, intended deployment and intended recipients, before initial deployment and before deploying a substantially modified version. The NSA would have 30 days to evaluate cyber and CBRN capabilities, extendable by 30 days with written notice, and 30 days after that to give the developer an unclassified report, optionally with a classified annex.
Deployment would be prohibited unless either the NSA has completed its evaluation and the developer has implemented its recommendations, or the NSA has not met its own deadlines.
The bill is one of several federal proposals made since the summer, including the FRONTIER Act and the AI Kill Switch Act. It would make release conditional on a government agency’s evaluation. A bill that is not enacted by the end of the Congress lapses.