Malaika Grounds Agentic Malware Analysis in Domain, Program Semantics, and Threat Intel

arXiv·medium signal

Malaika frames malware analysis as a grounded reasoning problem using three mechanisms — domain grounding to constrain hypotheses, semantics grounding to localize program evidence, and knowledge grounding for attribution — inside a multi-agent architecture combining analyst-inspired reasoning, tool-mediated evidence localization, and retrieval-based behavioral attribution. The authors report more precise and auditable conclusions than prior LLM-based and frontier agentic frameworks, with ablations confirming grounding is essential to reliable understanding.