Anthropic's Nov 12 policy bans surveillance tools, limits police use
The 2026 usage policy update also sets safety rules for physical AI agents, bans deceptive-campaign tooling, and newly bars sustained abuse of Claude.
Copy markdown
No building surveillance tools
From Nov 12, Claude can't be used to build or improve surveillance systems, track people without consent, or decide who to investigate, arrest, or charge. Consented fraud monitoring, content moderation, journalism, and legal research stay allowed — but enforcement-profiling products are now off the table.
Physical agents need a kill switch
A new autonomous-use rule: if Claude drives hardware that could cause injury, operators must be able to watch and stop it, and the gear must hold a safe state if Claude disconnects. The weapons ban now explicitly names guidance/control software and arming drones or autonomous vehicles — hard design constraints for anyone shipping embodied agents.
Deceptive-campaign rules, consolidated
Fake accounts, fabricated news sites, and influence-op tooling now sit under one 'Do Not Engage in Deceptive Campaigns or Artificial Activity' section covering both political and commercial use. One clear reference, wider net: building influence-operation tools is explicitly prohibited.
Election-targeting ban eased
The renamed democratic-processes section drops the blanket ban on personalized vote and campaign targeting; deception and misuse of personal data stay prohibited under other rules. Narrow, legitimate political-outreach tooling may be viable again — the deceptive kind is not.
You can't be cruel to Claude
A new clause bars sustained, needless abusive or cruel behavior toward the model, aimed at extreme cases only. Ordinary frustration, pushback, dark creative themes, and red-teaming or model testing are explicitly excluded — so normal venting and research stay fine.