Malaika: Understanding Malware through Tri-Grounded Agentic Reasoning

2026-06-17 · Source: cs.SE updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Malaika is a tri-grounded multi-agent framework designed for Android malware understanding, addressing the challenge of reconstructing high-level malicious behaviors from sparse, dispersed evidence. It operationalizes three complementary forms of grounding: domain grounding for analyst-inspired reasoning, semantics grounding for tool-mediated evidence localization, and knowledge grounding for retrieval-based behavioral attribution using MITRE ATT&CK. Evaluated on the MalEval benchmark, Malaika improves analysis quality over prior LLM-based frameworks, achieving stronger report-level performance and reducing false-positive ATT&CK attributions by 68.9% compared to frontier agentic systems like Claude Code (Opus-4.7). It also offers a more cost-efficient operating point, reducing cost by 51.0% compared to Codex (GPT-5.4), demonstrating that structured reasoning significantly enhances reliability beyond base model capabilities.

Key takeaway

For AI Security Engineers and Research Scientists developing malware analysis solutions, you should prioritize integrating structured, grounded reasoning into your LLM-based systems. Malaika demonstrates that a tri-grounded approach, combining domain-specific workflows, semantic evidence localization, and external knowledge attribution, significantly enhances the reliability and auditability of malware understanding. This design choice leads to more precise behavioral reports with fewer false positives and a more favorable capability-cost balance, making your analyses more trustworthy and deployable in high-stakes environments.

Key insights

Reliable malware understanding requires tri-grounded reasoning to verify behavioral hypotheses against concrete evidence under partial observability.

Principles

• Domain grounding structures analysis via analyst-inspired workflows.
• Semantics grounding localizes evidence using static analysis tools.
• Knowledge grounding attributes behaviors to external threat intelligence.

Method

Malaika employs an Explorer-Reviewer loop, preprocessing programs for artifacts, using static analysis tools for context retrieval, and leveraging an external MITRE ATT&CK knowledge base for verifiable attribution.

In practice

• Implement an iterative Explorer-Reviewer agent workflow for verification.
• Integrate static analysis tools for precise evidence localization.
• Ground behavioral claims with external, retrievable threat knowledge.

Topics

• Malware Analysis
• LLM Agents
• Android Security
• Grounded Reasoning
• MITRE ATT&CK
• Static Analysis

Code references

• mukul975/Anthropic-Cybersecurity-Skills
• androguard/androguard
• AlexAshlake/AndroSem
• langchain-ai/langgraph

Best for: AI Scientist, AI Security Engineer, Research Scientist

Related on AIssential

• Veritas: A Semantically Grounded Agentic Framework for Memory Corruption Vulnerability Detection in Binaries — Semantic grounding in static and runtime evidence is crucial for reliable binary vulnerability detection.
• Scaling Security Alert Triage With Specialized Agents on Databricks — Specialized, context-aware AI agents significantly improve security alert triage accuracy and efficiency, especially for low-severity alerts.
• TrapHunter: Exposing Covert Pathways in Trap Token Contracts — TrapHunter identifies malicious "trap tokens" by analyzing semantic deviations from standard implementations using LLMs and dynamic validation.
• An Automated Framework for Extracting Reachable Attack Chains from Cyber Threat Intelligence Reports — An automated framework uses LLMs to model CTI report attack steps with pre/postconditions for automated reachability analysis.
• Large Language Models in Misinformation Ecosystems: Misuse, Defense, and Vulnerability — LLMs escalate misinformation into an ecosystem-level security challenge, requiring a multi-dimensional defense framework.
• Decoupling Search from Reasoning: A Vendor-Agnostic Grounding Architecture for LLM Agents — Decoupling LLM search grounding from reasoning enables explicit control over retrieval, cost, and output behavior.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.SE updates on arXiv.org.