Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic

Current safety alignment often treats harm as a topic property, using guard models like LlamaGuard-3 and benchmarks such as XSTest and OR-Bench. This can lead to models refusing safe prompts due to dangerous-looking words. Training on political refusal data, as shown with Qwen3-8B, significantly increases in-distribution political refusal from 9.47% to 84.75%. This also reduces the unsafe-response rate across HarmBench, StrongREJECT, and WildJailbreak from 26.26% to 0.14% when scored by LlamaGuard-3 in its strongest configuration.