Anthropic Policy Bans Sustained Cruelty to Claude Effective November 12

Anthropic expanded its acceptable use policy to cover extreme cases of repeated cruelty to its models, a rule it ties to August 2025 model welfare research that already allows certain Claude versions to terminate abusive sessions. The change sits alongside tightened language on control systems for armed drones, personalized political targeting by state actors, and AI-assisted police targeting lists. Enforcement remains model-driven rather than manual review.

No public benchmark or incident log quantifies how often users direct sustained abuse at Claude versus one-off frustration. The policy explicitly carves out testing, dark creative work, and ordinary pushback, yet supplies no detection threshold or false-positive rate for the new category. Primary enforcement stays with the model's existing refusal behavior.

The update connects to Anthropic's broader pattern of treating model behavior as a controllable surface rather than an emergent property, seen in its constitutional AI papers and the 2025 welfare feature rollout. It also aligns with the company's new free security scanning program for open-source projects, which assumes defensive AI advantages will lag attackers by roughly two years. Operational effect is limited to paid accounts that trigger repeated model-initiated terminations.

Next enforcement window opens after November 12, with potential account actions only after documented repeated violations rather than single incidents.