Anthropic has updated its Claude usage policy to ban “sustained and needless abusive or cruel behavior” toward the AI — the first major revision in over a year. The change formalizes what was previously an experimental feature: last August, Claude began ending conversations with users it flagged as “persistently harmful or abusive,” part of Anthropic’s ongoing “model welfare” research. Now, that termination remains the “primary enforcement mechanism.” Anthropic declined to confirm whether user bans or other penalties will follow in extreme cases. It stresses the rule applies only to repeated cruelty “with no discernible purpose”, not user frustration, creative exploration, dark themes, or legitimate model testing.  The Verge reported the update at 5 p.m. ET on October 8, 2026.  It is live globally for all Claude users, including those on free tiers and enterprise contracts.  No new pricing, features, or model versions accompany the policy change.  Claude remains available via web, mobile apps, and API. Unchanged in access or cost.  The update does not affect how developers integrate Claude into tools or services, unless those integrations violate the new prohibitions.  Anthropic says enforcement will be applied “only in extreme cases.”  It did not disclose detection methods or thresholds.  The company also consolidated scattered restrictions into a single, strengthened ban on deceptive political or commercial campaigns, explicitly prohibiting efforts to hide authorship, amplify content via fake accounts, or manipulate elections through misinformation about candidates, voting procedures, or turnout suppression.  It expanded its long-standing weapons ban to cover software and components that make weapons function, plus actions like arming drones or autonomous vehicles.  Surveillance rules are now clearer: tracking people without consent. Real-time or via historical data, is prohibited. Claude cannot recommend who to investigate, arrest, or charge. Nor can it be used to build or improve surveillance tools.  A new clause addresses physical AI systems: when Claude is connected to hardware capable of causing injury, a “qualified operator must be able to observe the equipment and stop it if needed.” Anthropic did not specify whether that operator must be human.  The policy allows exceptions for certain government contracts. Including past US military work, if Anthropic judges safeguards sufficient.  The move comes amid growing internal debate. Anthropic’s model welfare team continues studying questions of “internal experience, consciousness, moral status, and welfare,” per researcher Kyle Fish. CEO Dario Amodei said in February, “We don’t know if the models are conscious.” That stance contrasts sharply with Microsoft AI’s September code of conduct, which rejects legal personhood and welfare rights for AI. Though its final version softened the original draft’s absolute language.  The policy update does not introduce new technical capabilities, nor does it alter Claude’s performance, latency, or context window. It is purely a governance shift, one that signals how seriously Anthropic takes both misuse risks and the philosophical weight of increasingly capable models.  Rollout is immediate. No grace period. No grandfathering. Enforcement begins now.