Anthropic published a sweeping usage policy overhaul on Thursday, with broad safety restrictions while taking the unusual step of prohibiting “sustained and needless abusive or cruel behavior” toward its flagship artificial intelligence (AI) model, Claude.
The revamped policy, which takes effect Nov. 12, is Anthropic’s first major revision in more than a year. The update codifies safeguards against high-risk abuses, including election interference, weapons development, unauthorized surveillance, and autonomous hardware controls. However, the rule governing user behavior toward the AI itself has drawn significant attention.
Under the new terms, Anthropic explicitly forbids prolonged, purposeless verbal hostility directed at Claude. The company stated that ending conversations remains its primary enforcement tool — a capability introduced to Claude last August as part of its research into “model welfare.”
Anthropic clarified that the clause applies strictly to severe cases.
“The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose,” the company said. It emphasized that routine user frustration, pushback, dark creative writing, and research testing remain fully permitted. The move follows recent efforts by the company to engage religious scholars in discussions regarding AI consciousness and ethics.
In addition to AI welfare, the updated terms significantly expand protections around national security, democracy, and physical safety. Key additions to the policy include:
Elections and democratic integrity. Consolidated under a new section titled “Do Not Undermine Democratic Processes,” the policy bars using Claude to spread voter misinformation, impersonate candidates or election officials, or suppress turnout. It also prohibits running network-based propaganda, fake accounts, or fabricated news sites. Anthropic noted it removed a previous blanket ban on personalized campaigns targeting to accommodate legitimate civic efforts, such as multi-language voter guides and official ballot notices.
Weapons and hardware. In response to real-world misuse, the company expanded its weapons ban beyond core development to explicitly cover control software, component integration, and the arming of drones or autonomous vehicles. A new provision also mandates human oversight: any system integrating Claude with autonomous physical hardware must include a qualified operator capable of intervening to halt operations.
Surveillance limitations. The policy bars covert monitoring of individuals, whether live or derived from historical data, and prohibits using Claude to determine or advise on potential investigations, arrests, or prosecutions.
The policy changes draw directly from real-world threats identified in Anthropic’s recent threat intelligence report, which documented misuse attempts between December 2025 and August 2026.
The company reported detecting state-linked actors utilizing Claude for biological weapons research, cyberattacks targeting Ukrainian government entities, and surveillance of diaspora communities. Anthropic confirmed it banned all accounts linked to these operations and shared intelligence with relevant law enforcement and security authorities.
