Anthropic updates usage policy, added rules on model abuse and deception

Anthropic released its 2026 Annual Usage Policy Update on October 8, 2026, solidifying its rules against deceptive campaigns, adding a ban on abusive behavior against its models, and clarifying requirements around weapons, surveillance, and high-risk usage. The updated policy will take effect November 12, 2026.

Anthropic said most of the changes are intended to clarify existing rules. The company said the review responds to a year in which Claude has undertaken more extensive and independent work, incorporates customer feedback and is based on patterns of misuse in influence operations, weapons development and surveillance documented in its September 2026 threat intelligence report.

The Usage Policy, also called the Acceptable Use Policy, applies to anyone who may submit input to Anthropic’s products or services, including through authorized resellers or passthrough access. It is organized into universal use standards applicable to all users, high-risk use case requirements for specific consumer-facing uses with high risk of harm, and additional use-case guidance covering consumer-facing chatbots, child-facing products, use of Model Context Protocol agents, and servers. Anthropic’s safeguards team implements detection and monitoring to enforce the policy; violations may result in limiting, suspending, or terminating access, and Anthropic may block or modify model outputs when inputs violate the policy.

Deceptive campaigns and electoral rules

Anthropic said it has seen state media, government propaganda offices and commercial companies use Claude to operate networks of fake accounts and made-up news sites. Such activity was already banned, but the restrictions were scattered across sections related to elections, fraud, privacy and disinformation. The update consolidates these in a new section titled Do Not Engage in Deceptive Campaigns or Artificial Activities, which applies to deceptive activities of any kind, whether political or commercial. It covers efforts to obscure who is behind a message, the amplification of content through fake accounts or posts, and the building of tools and infrastructure to conduct influence operations campaigns.

The Elections Section has been renamed Do Not Undermine Democratic Processes and is specifically focused on prohibiting the use of Claude to deceive voters or disrupt elections, including spreading false information about candidates or how to vote, impersonating candidates or election officials, and suppressing voter turnout.

Anthropic has also removed the general ban on personalized voting and campaign targeting. The company said the rule covers legitimate civic work, such as nonprofits using Claude to write voter information in other languages and election officials sending out election correction notices. Targeting based on deception or the misuse of voters’ personal data remains prohibited in the sections relating to deceptive campaigns, surveillance and privacy. The current policy text, which will remain in effect until November 12, 2026, still lists personalized voting and campaign targeting among the prohibited political activities.

Weapons, surveillance and law enforcement

Anthropic said it has observed attempts to use its models to create guidance and control software for weapons. The updated section on weapons makes explicit that the bans include the software and components that operate the weapons, as well as actions such as arming drones and other autonomous vehicles. Anthropic said the changes reflect how it applied previous policy.

The rewritten section on surveillance and criminal justice prohibits tracking people without their consent, either in real time or through analysis of previously collected data. Claude may not be used to decide or advise who to investigate, arrest or charge in a police or criminal justice process, nor to create or improve tools designed for surveillance. Anthropic said its September report documented a growing use of artificial intelligence to build surveillance systems that identify and track political dissidents, and that changes in surveillance, like changes in weapons, make existing enforcement practices more explicit rather than changing them. The section also specifies the uses that remain permitted, including tracking that people have consented to, such as fraud monitoring, content moderation, journalism and legal research.

High-risk use cases and hardware controls

When Anthropic products are used in ways that may affect someone’s health, legal rights, finances, livelihood, or access to essential services, the policy requires that a qualified person be in the loop – someone with the authority to review and, if necessary, modify Claude’s recommendations – and requires that the affected individual be told that artificial intelligence was used. Anthropic said these requirements have not changed, but it has rewritten the section to more directly answer a frequently asked user question: It now lists which types of recommendations are covered and which are not.

Following the launch of Anthropic’s Model Hardware Standard research preview, the section adds requirements for models attached to hardware that take autonomous physical actions and may be capable of causing injury. A qualified operator must be able to observe the equipment and stop it if necessary, and the equipment must be able to maintain a safe state if Claude is disconnected.

Offensive behavior towards models

The update adds a ban on prolonged and unnecessary abusive or cruel behavior towards Anthropic models. The company said the rule is intended to apply only in extreme cases, where users repeatedly act cruelly towards models without any discernible purpose, and does not apply to ordinary user frustration, holdouts, dark creative themes, or model testing and searches.

The ban ties to a feature introduced by Anthropic in 2025. In an August 15, 2025 research post, the company said it has given Claude Opus 4 and 4.1 the ability to end conversations in its consumer chat interfaces, intended for “rare and extreme cases of persistently malicious or abusive user interactions.” Anthropic said the feature was developed primarily as part of its exploratory work on the potential well-being of AI, with broader relevance to model alignment and guarantees.

According to that post, Claude is advised to use the ability only as a last resort, when multiple redirect attempts have failed or a user explicitly asks to end a chat, and not when a user could be at imminent risk of harming themselves or others. When a conversation ends, the user cannot post new messages within it, but other conversations on the account are not affected, and previous messages can be edited and retried to create new branches. Anthropic said this capability, available on Claude.ai and Claude Code, will remain the primary enforcement mechanism for the new abuse ban.

Clarification on supported regions

The update also addresses the Supported Regions Policy, which Anthropic strengthened in 2025 for businesses majority-owned by entities located in unsupported regions. Anthropic said the clarified page explains how it enforces the policy: Use of Claude is prohibited by persons physically located in an unsupported region, by entities incorporated or headquartered in one, and by entities majority owned or controlled by persons or entities in those regions.

Anthropic said it will continue to update the usage policy as Claude’s capabilities and risks change, seeking input from policymakers, subject matter experts, civil society and people who use its products. The updated policy will take effect November 12, 2026.