Anthropic published a new Usage Policy for Claude on Thursday, TechCrunch reported. Anthropic said it takes effect November 12. Two additions are new. One is a set of controls for Claude connected to hardware. The other is a ban on sustained, pointless cruelty toward its models. Most other changes clarify existing rules, and one rule was dropped. Anthropic removed its blanket ban on personalized vote targeting. It said the ban caught civic work, such as voter notices in other languages. Deceptive targeting and misuse of voters' personal data stay barred under other sections.
Anthropic said its weapons and surveillance changes only make existing enforcement explicit. For hardware that could cause injury, a qualified person must be able to see the equipment and halt it. The equipment must also hold a safe state if Claude disconnects. Anthropic did not say how it will judge extreme abuse or what penalties follow. It said Claude ending abusive chats remains the main enforcement tool.
Usage policy is turning into an engineering requirement for agents. A stop control and a safe state treat the link to the model as something that can fail. Teams using Claude can check three things. Does any hardware setup have both controls? Do high-risk uses keep a human reviewer and an AI notice? Do campaign tools fit the new deceptive campaigns section?