Safety guardrails in Code Large Language Models (Code LLMs) are designed to prevent the generation of malicious code, yet their robustness under linguistically diverse inputs remains largely unexplored. To address this important safety issue, we present MALICE (Malicious Adversarial LLM Inputs for Code Generation), a large-scale adversarial benchmark that exposes a critical blind spot at the intersection of multilingual malicious prompts and code generation safety. Our approach converts malicious English prompts into code-mixed and transliterated variants through sentence-level translation into low-resource languages while preserving programming keywords intact. The resulting dataset comprises approximately 250,000 adversarial prompts spanning 18 languages at various resource levels, derived from nine existing malicious code benchmarks. We perform a comprehensive evaluation across nine models (including Claude 4.6 and GPT-5.2) and safety tools, which reveals concerning vulnerabilities: code-mixed malicious prompts reduce detection accuracy by over 0.70 absolute points for the strongest classifier, while even the most safety-aligned models see their malicious code generation rates increase threefold. These results demonstrate that current safety alignment fails to generalize beyond English inputs, underscoring an urgent need for multilingual defenses in Code LLMs.