Multilingual Evaluation of Malware-Generation Guardrails

Extend CS-Guard to determine how guardrails for code-generation security perform against malicious prompts in multilingual settings, thereby characterizing whether language-dependent vulnerabilities observed in general LLM safety evaluations also arise in malware-generation scenarios.

Background

CS-Guard evaluates only English-language data. The paper notes that prior research has identified weaknesses in LLM defenses against malicious prompts in multilingual settings, but that multilingual behavior has not been examined specifically for malware generation. The authors therefore identify a concrete unresolved research direction: evaluating code-generation security guardrails across languages, which would require substantial additional data collection and experimentation.