Anthropic updates its Usage Policy: sustained abuse of Claude banned, autonomous physical actions need a human who can stop them; effective Nov 12

Why it matters

【Key points】

Anthropic published its annual Usage Policy update on Oct 8, effective Nov 12. Most changes clarify existing rules and add examples for Claude's longer, more independent work.

Main changes: rules scattered across elections, fraud, privacy and disinformation are merged into a new section on deceptive campaigns and artificial activity, political or commercial. The elections section now targets deceiving voters and disrupting elections, and the blanket ban on personalized voter targeting is dropped. Weapons rules now name guidance and control software and arming drones and other autonomous vehicles. Surveillance rules bar tracking people without consent and using Claude to decide or recommend who to investigate, arrest or charge. High-risk uses still need a qualified human in the loop and disclosure; new requirements say that when Claude drives hardware that acts autonomously and could injure, an operator must be able to observe and stop it, and the equipment must hold a safe state if Claude disconnects. A new rule bans sustained, needless abuse of the models, for extreme cases only; ordinary frustration, dark creative themes and model testing are excluded, and ending the conversation stays the main enforcement. The Supported Regions page now spells out enforcement by location, incorporation and majority ownership.

【Verifiable】

Anthropic: https://www.anthropic.com/news/2026-usage-policy-update

【Our read】

1. The abuse ban drew the attention, but its bite is small: enforcement is still ending the chat. Its significance is putting model welfare into terms users must accept, which other labs will now be asked about.

2. The physical-actions rule matters more. As Claude reaches hardware, "a human can stop it and it fails safe" becomes the deployer's compliance duty.

3. The tighter surveillance and law-enforcement language, especially not choosing who to investigate, will shape how government and security customers can use Claude.