Large language models are entering classrooms and homes, yet the moderation tools that guard these interactions are aligned to adult users. We show that this produces a dangerous alignment gap: on our benchmark, general-purpose safeguards such as the OpenAI Moderation API and base Llama Guard 3 detect as little as 30–38% of child-directed harms—predatory grooming, emotional manipulation, encouragement of risky behaviour—because these harms carry no profanity and violate no adult-calibrated policy. Qualitatively, these tools pass textbook grooming as safe; we conclude that, in their current form, they are not on their own an adequate safeguard for children’s conversations with chatbots. To close the gap we (i) define a child-centric risk taxonomy of 27 categories across 7 domains, differentiated by developmental stage (ages 6–12 and 13–17); (ii) build a synthetic pipeline producing 83,338 single-message and multi-turn child–AI interactions, filtered to 79,193 by a reward-model quality mechanism; and (iii) fully fine-tune an encoder (ModernBERT-large) and LoRA-adapt generative safeguards (Llama Guard 3, Aegis-Defensive). Fine-tuned Llama Guard 3 8B reaches 0.996 macro-F1 and 0.994 harmful recall, cutting false negatives to near zero across all 27 categories. We give an honest account of the limits of fully synthetic evaluation—strong lexical baselines (BM25, 0.90 macro-F1) reveal distributional artefacts—and argue that generative safeguards remain the more robust choice for deployment. We release the taxonomy, dataset, and fine-tuned models, and outline planned validation of the safeguard inside a non-profit educational chatbot already used by tens of thousands of children, plus a free public moderation service for benchmarking child-safety.