Buried in a routine policy rewrite, between a ban on arming drones and a ban on surveillance tooling, is a rule about how you are allowed to treat the model itself.

The rewritten policy takes effect 12 November. The new prohibitions are mostly about physical and civic harm: weapons software and components, arming drones and autonomous vehicles, tracking people without consent, building surveillance tools, and using Claude to decide who to investigate, arrest or charge. Where Claude controls hardware that acts in the world, a qualified operator must be able to watch and stop it, and the equipment must hold a safe state if Claude disconnects. Then there is a clause with no precedent in a commercial AUP: a prohibition on sustained, needless cruelty toward the models, which Anthropic says targets extreme cases and not ordinary frustration, pushback, dark creative themes or testing. One rule went the other way: the blanket ban on personalised political targeting is lifted, so nonprofits and election officials can run voter information work.

Why this one is different

Usage policies forbid harm to people. This one also forbids a kind of conduct toward a product, which only makes sense if the company believes something about the product that it has not fully said. Anthropic frames it narrowly and declines to argue the philosophy, which is the careful choice; the clause exists either because the model's welfare might matter, or because what people practise on a model they practise. The other notable move is the loosening. Tightening weapons and surveillance while relaxing political targeting is not the direction a policy drifts by accident.

A commercial usage policy now has a cruelty clause.

How we got here

1. 9 Sep 2026Anthropic's alignment lead puts extinction risk above 10% as a researcher resigns.
2. 18 Sep 2026Anthropic names Accenture its first embedded evaluator.
3. 28 Sep 2026Three lab scientists sign the intelligence explosion paper.
4. 8 Oct 2026The usage policy adds weapons, surveillance and cruelty to the models.

What it does and does not mean

Anthropic says most of this makes existing enforcement explicit rather than changing it, which is the usual reading of a rewritten policy, and nothing here is enforceable by anyone but Anthropic. There is no stated test for sustained cruelty, no appeal described, and no figure for how often anything in this document has been enforced. The lifted political rule is the one with immediate practical effect, and it loosens a line the company drew when elections were the dominant worry. What the document records is a company writing down obligations in both directions at once, to the people its model could be used against, and to the model.