If you build on Claude, there's a new rulebook landing November 12. Anthropic folded its scattered election and disinformation rules into a single "deceptive campaigns" section covering fake accounts and fabricated news sites, political or commercial. The weapons language now reaches software and components, and bans using Claude to arm drones or to make biological and chemical agents more lethal or transmissible. A new surveillance clause rules out building tools that decide who to investigate, arrest, or charge, and anything that controls physical hardware now needs a qualified human who can watch it and hit stop. Two changes are worth a second look. Anthropic dropped its blanket ban on personalized vote and campaign targeting — deception and misuse of voter data stay prohibited under other sections, but the targeting prohibition itself is gone. And there's a narrow clause against sustained, needless abuse of the models themselves, which the company stresses is for extreme repeat cases, not ordinary frustration or dark creative work. Most of it reads as clarification after a year of Claude doing longer, more autonomous work. The previous policy holds until the 12th.