The machines are watching

Abstract

Algorithmically Generated Domains (AGDs) are integral to many modern malware campaigns, allowing adversaries to establish resilient command and control channels. While machine learning techniques are increasingly employed to detect AGDs, the potential of Large Language Models (LLMs) in this domain remains largely underexplored. In this paper, we examine the ability of nine commercial LLMs to identify malicious AGDs, without parameter tuning or domain-specific training. We evaluate zero-shot approaches and few-shot learning approaches, using minimal labeled examples and diverse datasets with multiple prompt strategies. Our results show that certain LLMs can achieve detection accuracy between 77.3% and 89.3%. In a 10-shot classification setting, the largest models excel at distinguishing between malware families, particularly those employing hash-based generation schemes, underscoring the promise of LLMs for advanced threat detection. However, significant limitations arise when these models encounter real-world DNS traffic. Performance degradation on benign but structurally suspect domains highlights the risk of false positives in operational environments. This shortcoming has real-world consequences for security practitioners, given the need to avoid erroneous domain blocking that disrupt legitimate services. Our findings underscore the practicality of LLM-driven AGD detection, while emphasizing key areas where future research is needed (such as more robust warning design and model refinement) to ensure reliability in production environments.