Anthropic bans sustained abuse of Claude from November 12. What counts, and what doesn’t

What changes on November 12

On October 8, Anthropic published a new version of its Usage Policy and said the update “takes effect on November 12.” Most of it is about deceptive campaigns, weapons, surveillance, elections and high-risk uses. One new line, under the heading “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct,” is about how you treat Claude. Users may not:

Engage in sustained and needless abusive or cruel behavior toward our models

Anthropic’s announcement narrows it: “The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose. It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.”

The rule drew a second round of attention on November 1, when CNN wrote about it. Anthropic declined CNN’s interview requests, but a spokesperson told CNN, in CNN’s words, that “the main enforcement mechanism will still be Claude ending the conversations, and that the prohibition is not for people expressing frustration, blunt feedback, or testing and research.” CNN reports that Mustafa Suleyman, who runs Microsoft AI, answered the policy with a melting-face emoji, and that Elon Musk called it “the right move,” writing: “Cruelty to something that believes it is experiencing pain is not ok.”

The policy follows your account, not the app. It “applies to anyone who submits inputs to Anthropic’s products or services,” naming Claude.ai, Claude Code, the API, cloud providers and “the end users of products or services integrating Claude.” That includes Claude Code run from an app on your Claude plan, Enjoy included.

What counts, and what doesn’t

Anthropic hasn’t defined “needless” or said where frustration ends and cruelty begins. Here is what it has said, and where:

What you do	Covered by the rule?	Where it says so
Repeated cruelty toward Claude “with no discernible purpose”	Yes	Anthropic’s announcement
Venting when Claude gets something wrong	No: “common versions of user frustration”	Anthropic’s announcement
Pushing back, disagreeing, blunt feedback	No	Anthropic’s announcement; its spokesperson to CNN
Fiction with dark themes	No	Anthropic’s announcement
Testing the model, or research	No	Anthropic’s announcement
Swearing at Claude Code	Doesn’t qualify for ending a session	Claude Code’s docs

The last row comes from Claude Code’s tools reference, which says of its session-ending tool: “General frustration, profanity, or a task going badly don’t qualify, and neither do requests for harmful content, which Claude declines instead of ending the session.” Requests for harmful content fall under the policy’s other rules, not this one.

What happens when Claude ends a conversation

Ending the conversation is the penalty Anthropic says it will mostly use, and it isn’t new. Claude has been able to end chats on claude.ai since August 2025, and Claude Code got an EndConversation tool in version 2.1.214, on July 18, 2026, according to its changelog.

On claude.ai, Anthropic says you “will no longer be able to send new messages in that conversation,” but your other conversations aren’t affected, you can start a new chat right away, and you can “edit and retry previous messages to create new branches of ended conversations.” Claude is told to use it only “as a last resort when multiple attempts at redirection have failed,” or when you ask it to end the chat, and never when someone “might be at imminent risk of harming themselves or others.”

In Claude Code, the tool is for “a last resort against sustained abusive input, after attempts to redirect the conversation have failed and after a clear warning in an earlier message,” or for when you ask to see it demonstrated and confirm. After that, the session locks, and prompts get “Claude ended this conversation. Start a new session (or /clear) to continue.” Only /clear, /resume, /help, /exit and /feedback still run. Resuming the session brings the lock back, but “the session’s history isn’t deleted.” A script that resumes an ended session with -p exits with code 1.

You can’t switch the tool off: deny rules, --disallowedTools and a --tools list don’t remove it “while any other tool remains.” It appears only:

• on Claude Opus 4.8, Sonnet 5, Fable 5, “or a later version of one of those families”;
• in an interactive terminal session, including an IDE’s built-in terminal. It isn’t in -p runs, the Agent SDK, the VS Code extension’s panel, GitHub Actions or cloud sessions;
• not through Amazon Bedrock, Claude Platform on AWS, Google Cloud’s Agent Platform, Microsoft Foundry or a cloud gateway.

The policy applies everywhere, whether or not the tool is there.

Can Anthropic ban you for it?

It can, though Anthropic says ending the conversation is how it will mainly enforce this rule. The Usage Policy’s enforcement clause covers every rule in it: “If we suspect that you may have violated our Usage Policy, we may warn you or throttle, limit, suspend, or terminate your access to our products and services.” For this one, Anthropic’s announcement says “Claude’s ability to end these interactions will remain the primary enforcement mechanism.” CNN’s reading is that “some users could, in extreme cases, be permanently banned.”

What Anthropic hasn’t said:

• what would take it past ending a chat, such as a number of ended conversations;
• whether people review ended conversations, and how it will judge “needless”;
• how the rule works on the API and other places where Claude has no way to end the conversation.

If your account is banned, Anthropic’s appeals page says to “log in with your banned account, and fill out the appeal form.” A Free, Pro, Max or Ultra account banned for breaking the Usage Policy can still log in to delete the account or export its data, though what it can export “may be restricted depending upon the Usage Policy violation.” If you think a warning was a mistake, Anthropic asks you to email usersafety@anthropic.com “with the details of your situation and your account information.”

Do ChatGPT, Gemini and Grok have a rule like this?

Not as we read their policies on November 2. OpenAI’s usage policies, Google’s Generative AI Prohibited Use Policy and the Acceptable Use Policy of SpaceXSI, the company behind Grok, all forbid using their services against other people. OpenAI’s rules out “threats, intimidation, harassment, or defamation,” Google’s “Harassment, bullying, intimidation, abuse, or the insulting of others,” and SpaceXSI’s “illegal, harmful, or abusive activities.” None has a rule about how you treat the model itself.

What to do

1. Vent if you need to. Frustration, pushback and swearing at a task that’s going badly are outside the rule, in Anthropic’s own words.
2. Take a warning as the signal. In Claude Code, Claude ends a session over abuse only after “a clear warning in an earlier message.” If Claude says it will end the conversation, what you send next decides it.
3. If a session ends, start another. Run /clear or start a new session in Claude Code; on claude.ai, start a new chat or edit an earlier message. In Claude Code, the session’s history stays. If you think Claude got it wrong, send /feedback in Claude Code, or use the thumbs or “Give feedback” button on claude.ai.
4. If you script Claude Code, check its exit code. Resuming an ended session with -p fails with code 1, so a script won’t mistake it for success.
5. Say what’s wrong with the work. It’s the quicker fix anyway. Jindong Wang, who runs an AI lab at William & Mary, told CNN that “how you talk to AI can affect the performance.” CNN adds that the evidence is mixed: Rosie Campbell of Eleos AI Research said some studies have found rude questions can improve results.

We’ll update this memo if Anthropic says more about how it enforces the rule after November 12.