Anthropic tightens Claude rules on deception, weapons, surveillance and abuse

The updated policy takes effect on November 12 and gives builders clearer limits on uses that involve elections, tracking people and weapon support software.

Anthropic has revised Claude’s Usage Policy to tighten rules around deceptive campaigns, election interference, weapons-related software, surveillance, high-risk deployments, and sustained abuse of the model, with the new version taking effect on November 12. For builders, the most practical changes are that Anthropic now groups fake-account and influence-operation activity into a single deceptive-activity ban, spells out election misuse more clearly, and extends weapons restrictions to software and components that make weapons function. The company also says Claude cannot be used to track people without consent, to choose who law enforcement investigates or charges, or to build or improve surveillance tools, while still allowing some uses such as fraud monitoring, journalism, and legal research. On the deployment side, Anthropic says high-risk uses must still include a qualified human reviewer and user notice, and hardware-connected systems must be observable, stoppable, and able to fail safely if Claude disconnects. It also added a narrow prohibition on sustained cruel behavior toward the model, aimed at repeated abusive interactions with no clear purpose.