Congress floats new AI oversight rules, and open weights draw fresh warnings

A bipartisan House bill would impose audits, incident reporting, and even criminal penalties on frontier AI developers, while a separate dispatch argues that publishing model weights makes dangerous systems impossible to recall.

Six House members from both parties introduced the FRONTIER Act, built from parts of an earlier draft framework and requiring safety frameworks, third-party audits, independent verification, and prompt reporting of critical safety incidents. (Joe)
The bill would create an Under Secretary for AI Security at Commerce, allow emergency orders halting a model that poses imminent catastrophic risk, and back it all with daily fines and possible prison time, a first for AI safety law. (Joe)
Gaps remain: nobody is required to license the independent verifiers the law depends on, funding for the federal evaluation agency and whistleblower protections were dropped, and the bill would override both state laws and other federal restrictions in the same areas. (Joe)
After a cyberattack, Hugging Face relied on a Chinese open model because US models' safeguards blocked the work, since attackers can pose as defenders to get help. (Donald)
Public weights cannot be taken back and their safety measures are easy to strip, which is why Yoshua Bengio urges sharing only models evaluated as below a risk threshold, though private labs have also lost control of their models. (Donald)