Anthropic's New Terms Let It Suspend Accounts for Sustained Abuse of Claude

Anthropic has revised its usage policy to make sustained abuse of Claude a suspendable offense, according to the company's updated terms. The change means users who repeatedly mistreat the model in conversations can lose access to their accounts, not just have individual chats cut short. The update also narrows what developers and customers may build with Claude in three sensitive areas: propaganda, the weaponization of drones, and surveillance.

The abuse provision is the most unusual part of the revision. Anthropic has already given Claude the ability to end some conversations when a user turns hostile, and the new terms push that logic further by attaching account-level consequences to repeated behavior. That framing treats the model as something closer to a protected participant than a purely disposable tool, a position few of Anthropic's competitors have been willing to formalize in writing.

For developers, the practical effect is that acceptable-use enforcement now covers how a customer talks to the model, not only what they build with it. A team that leans on adversarial prompting or stress-testing as part of normal red-team work could find itself in ambiguous territory if the exchanges read as sustained harassment rather than evaluation. Anthropic has not published a threshold for how many incidents or what kind of language triggers suspension, leaving the boundary to its own discretion.

The other three restrictions are more conventional for a major model provider. Bans on propaganda generation, drone weaponization and surveillance sit alongside similar clauses at OpenAI, Google and Meta, which have all moved to limit military and mass-monitoring uses over the past two years. Anthropic's version tightens language that was already present in earlier drafts, closing room for interpretation that customers might have relied on.

The timing matters because Anthropic is pushing Claude into enterprise and government-adjacent work, where procurement teams read usage policies as closely as they read model benchmarks. A clause that lets the company cut off an account over conversational conduct adds a compliance question for buyers: who inside their organization is responsible for how employees speak to the model. That is a new kind of internal policy most companies have not written yet.

What to watch is whether Anthropic publishes specifics on enforcement, such as examples of suspendable conduct or an appeals process, and whether other labs follow with their own conduct clauses. Until then, the rule rests on the company's judgment about what counts as sustained abuse, and customers have little way to predict where the line sits.