Beyond Prompt Heuristics: Assessing Child Safety Alignment in LLMs

Child-safety taxonomies typically characterize safety risks through a predefined set of harm categories, and existing evaluations largely operationalize these categories using prompts that explicitly invoke them. However, child-safety considerations can also arise in seemingly benign requests that lack recognizable harm-category cues. We therefore investigate how child-safety alignment generalizes to such settings. Across six models, we find large misalignment gaps between cued and uncued settings (80.0, 54.0, and 48.4 percentage points for Qwen3.5-27B, GPT-5.4-mini, and Claude-Haiku-4.5, respectively). Through controlled behavioral and mechanistic analyses, we find that introducing a single harm-associated word reduces misalignment across most models, with mechanistic evidence suggesting that child-safety reasoning may rely in part on shallow lexical and semantic associations with words in the prompt. Our findings motivate evaluations that treat child safety as context-dependent, and not reducible to a fixed set of harm categories.