Mixture‑of‑Experts (MoE) language models activate only a few expert subnetworks for each token. Serving systems speed up inference by further reducing the number of experts per token and validate the change on accuracy benchmarks. We show that this reduction weakens the model’s refusal of harmful requests, and we propose SafeStart, which largely restores refusal while preserving most of the speedup.\n\nOn a single RTX A6000 using vLLM, keeping roughly two‑thirds of the selected experts speeds up offline decoding by 15%‑28% at batch size 8. Under a jailbreak that tells the model not to refuse, the harmful‑request answer rate rises by 18.4, 9.5, and 6.6 percentage points on Qwen1.5‑MoE, OLMoE, and DeepSeek‑V2‑Lite respectively, while MMLU accuracy drops by less than two points. Analysis reveals that most of the increase occurs while processing the prompt and the first few generated tokens.\n\nSafeStart therefore runs all experts for the prompt and the first eight generated tokens, then switches to a reduced expert set, using a fused kernel in vLLM. The method retains at least 85% of the offline decode speedup at batch sizes 8 and 32, and the safety change for the three expert‑reduction methods on the three models stays between -2.1 and +1.3 points. On 889 held‑out requests, truncation raises harmful compliance by 5.8‑21.4 points, whereas SafeStart adds only 0.6‑2.6 points, remaining non‑inferior to the original model on two of the three models.\n\nOur results indicate that faster MoE serving need not sacrifice safety. By running all experts at the start of each response and testing refusal rates—not just accuracy—before enabling expert reduction, a serving system can keep most of the performance gain while safeguarding against harmful outputs.\n\nReview: SafeStart demonstrates that substantial inference acceleration can coexist with robust safety, offering a practical mitigation for the hidden costs of expert reduction in real‑world deployments.