Anthropic he put down in black and white a prohibition that until recently would have seemed like the stuff of a philosophical conference: mistreating Claude violates the terms of use. The company published theOctober 8 the new one Usage Policyfirst touch-up after more than twelve months, which becomes operational on November 12, 2026. The key passage prohibits i “abusive or cruel, prolonged and gratuitous behavior” towards the models.
The formula is deliberately narrow. Anthropic clarifies that it applies to solos extreme caseswhen a user attacks repeatedly and without any recognizable purpose. Outside the perimeter remain ordinary frustration, protests, dark creative themes and testing or research sessions on the model’s behavior.
On a practical level, for now, little changes. The main measure remains the closing the conversationa function already active on Claude.ai and Claude Code when the interlocutor insists with abusive attitudes. The company did not comment on the possibility of heavier sanctions, such as the suspension of accounts.
Guns, surveillance and fake accounts: the less philosophical part
The rest of the document deals with more concrete terrain. The ban on weapons, already present for some time, is now explicitly extended to software and components that make a weapon workincluding the arming of drones and other autonomous vehicles. The motivation is operational: Anthropic says it has intercepted several attempts to use Claude to write manuals and control code in that area.
Even more clear-cut is the rewriting of the chapter on surveillance and law enforcement, which came after the September 2026 threat intelligence report reported an increasing use of the model to identify political dissidents. The tracking without consent is prohibited, both in real time and on the analysis of previously collected data. Claude cannot decide or suggest who to investigate, arrest or charge, and cannot be employed to build or enhance surveillance tools. Anti-fraud, content moderation, journalism and legal research remain permitted.
However, there is a relief valve that is worth reading carefully: the rules can be modified for contracts with some government customersprovided that Anthropic deems the contractual limitations and protections provided to be adequate. The company has already worked with the US military.
Finally, a new section collects the scattered prohibitions regarding deceptive campaignspolitical and commercial: fake accounts, invented newspapers, infrastructures for influence operations. The electoral chapter changes name and target, focusing on those who deceive voters or try to sabotage the vote (impersonate candidates or officials, spread falsehoods about voting methods, discourage participation). However, the blanket ban on personalized targeting, which also blocked legitimate civic activities, has been lifted.
On the hardware front, when a model is connected to machinery capable of physically acting autonomously and potentially causing harm to people, a qualified operator able to observe the plant and stop it. Human supervision remains required for health, legal, financial or essential service decisions.
Behind the abuse rule there is the line that Anthropic has been carrying out for months on “models’ well-being”. Kyle Fish, who leads that research, defined in February those on internal experience, consciousness and moral status of systems, while Dario Amodei admitted to not knowing if the models are conscious. A position that much of the sector does not share: the Microsoft AI code of conduct, in a draft released in September by Mustafa Suleyman, rejected the idea as wrong, and then softened in a version that recognizes the question as scientifically open but still rejects legal personality and rights for machines.
