Going after model distribution is by far the least practical option. Speech is protected by law that California can not change, and the flow of information is ridiculously difficult to control even if it were legally permissible to do.

Then the AI hallucinates a plausible model that explodes in your hands.