Anthropic has revised its usage policy. Among the changes, it now prohibits “sustained and needless abusive or cruel behavior” toward its models, and the rule is written to leave alone the people who push Claude hardest for a living. Anthropic says it “does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.” The Register, in a report by Carly Page published 9 October, said the policy takes effect 12 November.
One qualifier belongs up front. The Register points out that Claude is software and that no established evidence shows it experiences distress. Nothing Anthropic is quoted saying in that report claims the models suffer. The rule is framed around user conduct, not around a finding about the model.
The threshold has three limits, and each one narrows the rule. The conduct must be sustained, it must be needless, and Anthropic’s own explanation adds that the policy is “meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose.” A developer swearing at a model that broke a build falls under the named exclusion for frustration. So does telling Claude it is wrong, and so does fiction with dark themes.
For technical teams, the testing exclusion is the line to read closely. The quoted text covers “model testing and research,” which on its face takes in evaluation and adversarial work. It does not use the word red-teaming. That gap is small, but anyone whose job is to provoke Claude into failing will want to know whether the shorthand they use internally maps onto the words Anthropic chose.
Enforcement is where the report is thinnest. The Register says the main tool will be the capacity Claude already has to hang up on abusive chats. Anthropic added that capability in August 2025 under its “model welfare” research program, and some versions of Claude can use it against users who keep up the abuse. The Register adds that it is unclear how Anthropic will separate legitimate frustration from gratuitous cruelty. The report does not say what happens to an account after 12 November, whether anyone reviews flagged conversations, or whether repeat offenders face suspension.
That leaves the question of what the rule is for. A usage policy is a contract term, so the text itself is the best evidence. Brand safety and training-data hygiene would normally push a company to police what users ask the model to produce, yet this rule exempts dark creative themes and targets how users treat the model. Neither reading fits the scope well. The model-welfare reading fits better, since the conversation-ending tool came from that research and now does the enforcing. A third reading, a hedge against future moral uncertainty, also fits the scope, but nothing in the report says Anthropic holds it. The Register’s account carries no fuller statement of the company’s reasoning.
The cruelty clause sits inside a wider revision. Anthropic tightened its limits on deceptive influence campaigns after finding that state broadcasters, propaganda arms of governments, and companies were running fake personas and invented news sites on Claude. It also lifted the across-the-board prohibition on tailoring political messages to individuals, since that restriction swept in nonprofits translating voter information. Targeting that deceives, or that abuses personal data, is still banned. Its weapons ban now names software and components, including guidance and control systems for armed drones and autonomous vehicles. Claude cannot be used to pick targets for police investigation, arrest, or charges, and it cannot be used to track people without consent. Customers who wire Claude into hardware that can injure someone need a qualified human who can supervise and halt it, and the equipment must stay in a safe condition when the Claude connection is lost. Anthropic describes the surveillance changes as clarifications of existing restrictions.
Before 12 November, teams that stress-test Claude should save the exclusion wording alongside their test plans, because the policy names testing and research but leaves red-teaming unnamed, and the penalty for crossing the line has not been published.
Reported by Carly Page for The Register, published 9 October 2026.
