Large Language Models (LLMs) can be prompted to generate malicious code. However, most existing studies focus on examples, qualitative analyses, or isolated payloads, leaving open the question of how LLM-generated malware compares with real-world malware. This work presents an end-to-end methodology and the first measurement-based study that systematically compares malware generated by censored and uncensored LLMs against 119 real-world malware samples from multiple families and categories, estimating how safety-alignment affects malicious capabilities. We assess the malware’s quality using metrics such as Tactics, Techniques, and Procedures (TTPs) coverage and resistance to static detection, which reflect how malware is analyzed in practice. Our findings suggest that LLMs can produce malicious code similar to real-world malware, replicating over 55
Acoustic side-channel attacks (ASCAs) recover typed text by exploiting the unique physical sounds of individual keystrokes. Standard defenses rely on degrading raw character-recognition accuracy, but recent work shows that downstream LLMs can easily correct heavily corrupted character streams, making standard evasion techniques ineffective for user privacy. This creates an important security paradox: if a downstream semantic corrector can salvage garbled predictions, is leading the raw acoustic classifier astray really protecting the user? We disagree. We present a new psychoacoustic adversarial framework for generating per-sample perturbations on live keystroke audio, bounded by an energy-adaptive spectral constraint. Our defense uses an untargeted Projected Gradient Descent (PGD) optimizer that is constrained at each iteration by a dynamic auditory masking threshold based on the frequency magnitude spectrum of the clean keystroke. We evaluate this framework on a state-of-the-art ResNet-34 acoustic-to-text classifier (adapted from the ASCA-Net architecture) on four different hardware keyboard platforms (HP, MSI, MacBook, and Lenovo) with two representative perturbation budgets ( epsilon=0.15andepsilon=0.50 ).Our adversarial keystrokes achieve a 92.4
Visual malware family classification provides a practical defensive route for recognizing malware variants without relying on full manual reverse engineering. However, existing image-based methods often emphasize spatial texture learning, while the structural cues contained in frequency-domain subbands are either ignored or compressed into low-dimensional statistics. This weakens their ability to distinguish visually similar or obfuscated malware families under lightweight deployment constraints. To address this problem, this paper proposes a Dual-Branch Wavelet Gated Fusion Network (DWGF-Net) for visual malware family classification. The method converts malware binaries into images, constructs four-channel level-2 discrete wavelet transform (DWT) subband maps from grayscale malware images, and processes these subbands with an independent wavelet branch. A spatial branch captures visual texture semantics, while LayerNorm-based alignment and a spatially guided gate adaptively fuse spatial and wavelet representations at the sample level. An auxiliary spatial-branch loss is introduced to stabilize the gating prior during training. Experiments on Malevis, BIG2015, and Malimg show that DWGF-Net achieves competitive defensive classification performance with only 0.649M parameters and low inference latency. The results further show that the contribution of wavelet-domain evidence varies across malware-image distributions, providing useful insight for deploying visual malware classifiers under lightweight constraints.
The growing prevalence of polymorphic, metamorphic and obfuscated malware has reduced the effectiveness of signature-based detection systems and created significant challenges for proactive malware defence. Existing learning-based malware classifiers typically rely on either API call sequences or control flow graph (CFG) analysis independently, limiting their ability to capture complementary behavioural and structural characteristics. Moreover, most multimodal approaches perform late-stage feature fusion and do not explicitly model semantic relationships between program behaviour and execution topology, while adversarial robustness is generally evaluated after training rather than incorporated into the learning process. This paper proposes MalGraphNet, an adversarially robust heterogeneous graph neural network framework for behaviour-based malware classification that jointly models API behavioural semantics and CFG topology within a unified heterogeneous graph representation. The framework introduces a heterogeneous malware behaviour graph linking API invocation, CFG basic-block and function-level nodes through multiple relation types, including cross-modal semantic bridge edges. A Relation-Aware Graph Attention mechanism performs heterogeneous message passing, while a Cross-Modal Bilinear Attention module captures fine-grained interactions between behavioural and structural representations. To improve resilience against malware obfuscation and adversarial manipulation, the proposed framework incorporates Adversarial Consistency Learning and Topology Preservation Regularization as intrinsic training objectives. Experimental evaluation on the EMBER and MalVis-2022 datasets demonstrates that MalGraphNet achieves 98.7
SocGholish is a malware dissemination framework that has been in operation since at least 2018, using compromised websites and fake browser update notifications to distribute malware to users. There is very limited research literature on SocGholish, with most existing research coming from threat reports by vendors rather than in-depth research articles or books. This paper presents a static analysis of 160 SocGholish samples from MalwareBazaar, including JavaScript loaders, PE executables, PowerShell scripts, and HTML decoy pages. A Python-based framework was developed to analyze SocGholish, which extracts over 70 features from each sample, including Shannon entropy, obfuscation, embedded network artifacts, and API calls. JavaScript loaders had a mean entropy of 5.36, with obfuscation patterns consistent among samples. PE payloads had entropy up to 8.0, consistent with packing or encryption. Among all samples, 54.4 p = 0.10 , ROC AUC 0.55), and on a size-matched subset they sit only 0.07 bits higher (AUC 0.71); this supports the reading that the loader stage is engineered to resemble benign content. Second, the framework’s own six-component confidence score fails on this task (ROC AUC 0.30, below chance), which reflects a bias in the scorer toward large bundled libraries rather than evasion by the samples, and which bounds what this particular feature set can achieve. These results offer a structured and reproducible data set and methodology to analyze fake browser update campaigns, and we discuss the implications of this for both detection engineering and threat intelligence.
