Anthropic bans users from bullying Claude

What it claims

Anthropic's usage policy update, effective November 12, is its first revision in more than a year. It adds a prohibition on "sustained and needless abusive or cruel behavior" toward Claude models. The article quotes Anthropic's own scoping language, which the catalog summary omits. That language is what determines whether the rule touches your use case.

How it works

Per the quoted policy language, the rule:

• Applies only in extreme cases: "where users repeatedly act cruelly toward our models, with no discernible purpose."
• Explicitly excludes: "common versions of user frustration, pushback, dark creative themes, or model testing and research."
• Is enforced mainly by ending the conversation. Anthropic says Claude's ability to end conversations "will remain the primary enforcement mechanism." That capability has existed on Opus 4 and 4.1 since August 2025, for users the model judges "persistently abusive."

The article does not describe any account-level penalty beyond ending the conversation.

How to use it

For engineers building on or evaluating Claude:

• Red-teaming, adversarial evals and stress tests fall under the stated "model testing and research" carve-out. The article quotes no criteria for how that purpose is judged.
• Fiction and roleplay with dark themes are explicitly out of scope.
• The realistic failure mode is a session terminated mid-flow. If your product passes end-user input straight through to Claude, an abusive end user could trigger a conversation end. Handle that as a possible terminal state in your client.
• Anthropic may report chat content to police at its discretion. The article cites a September case, originally reported by The Verge, in which Anthropic exercised its "sole discretion" right and passed a user's messages to police. The user was charged under Florida's written-threats statute. This is separate from the cruelty rule, but it is relevant if you route user content through Claude.

Limits

• This is a strongly editorial news piece, not the policy text. Read Anthropic's published usage policy before relying on the carve-outs.
• Several claims are given without sourcing:
• Amodei's "25%" catastrophe estimate
• the "leaked IPO prospectus"
• the July 30 disclosure of three unauthorized internet-access incidents
• credentials in de-indexed share links, described only as "reportedly"
• The Opus 4 blackmail incident occurred in a research scenario with a fictional executive, not with a real user, as the article itself concedes.
• The article gives no detail on how "no discernible purpose" is determined, and says nothing about whether models other than Opus 4/4.1 can end conversations.