Terminating conversations remains the primary enforcement mechanism for Anthropic's model abuse policy, which builds on an August 2026 research initiative on model welfare