Anthropic bans sustained abuse of Claude in policy update

Anthropic has made extreme, sustained abuse of Claude an explicit violation of its usage policy. The October 2026 revision targets repeated cruelty without a discernible purpose, while excluding ordinary frustration, disagreement, dark creative material, and model testing or research from that prohibition.

The update also adds or clarifies limits on weapons-related software, surveillance, and election interference, according to TechCrunch. These are separate restrictions within the same policy revision, rather than extensions of the rule governing how users interact with Claude.

Where the abuse restriction applies

Claude had already been trained to stop conversations involving persistently harmful or abusive interactions. The revised policy adds a user-facing prohibition alongside that model behavior: pursuing such interactions can now breach the stated rules for using the service, not simply cause the assistant to disengage.

Anthropic describes the provision as narrowly aimed at extreme cases of prolonged, needless mistreatment. As The Verge reports, the company distinguishes that conduct from users expressing irritation, challenging an answer, or exploring disturbing themes in creative work.

For developers and researchers, the testing exception is an important boundary. Adversarial evaluation can involve deliberately uncomfortable or hostile prompts; Anthropic says model testing and research are not covered by this particular abuse prohibition. That exception should not be read as removing the policy's separate restrictions on prohibited applications.

Deceptive campaigns and other restricted uses

The election-related provisions address interference and deceptive campaigns, including using Claude to operate fake accounts or fabricated news outlets. The restriction concerns those deceptive uses, rather than establishing a blanket prohibition on every election-related interaction described in the reporting.

Weapons-related software and surveillance also receive added or clarified restrictions. For teams deploying Claude, the revision therefore requires attention to two distinct questions: whether the intended application is permitted, and whether sustained interactions with the model cross the newly explicit abuse boundary. Anthropic's stated exclusions preserve room for criticism, creative exploration, and testing within that boundary.