Nowadays, The Universal Serial Bus (USB) is one of the most adopted communication standards. However, the ubiquity of this technology has attracted the interest of attackers. This situation is alarming, considering that the USB protocol has penetrated even into critical infrastructures. Unfortunately, the majority of the contemporary security detection and prevention mechanisms against USB-specific attacks work at the application layer of the USB protocol stack and, therefore, can only provide partial protection, assuming that the host is not itself compromised. Toward this end, we propose a USB authentication system designed to identify (and possibly block) heterogeneous USB-based attacks directly from the physical layer. Empirical observations demonstrate that any extraneous/malicious activity initiated by malicious/compromised USB peripherals tends to consume additional electrical power. Driven by this observation, our proposed solution is based on the analysis of the USB power consumption patterns. Valuable power readings can easily be obtained directly by the power lines of the USB connector with low-cost, off-the-shelf equipment. Our experiments demonstrate the ability to effectively distinguish benign from malicious USB devices, as well as USB peripherals from each other, relying on the power side channel. At the core of our analysis lies an Autoencoder model that handles the feature extraction process; this process is paired with a long short-term memory (LSTM) and a convolutional neural network (CNN) model for detecting malicious peripherals. We meticulously evaluated the effectiveness of our approach and compared its effectiveness against various other shallow machine learning (ML) methods. The results indicate that the proposed scheme can identify USB devices as benign or malicious/counterfeit with a perfect F1-score.

Rigorous testing methods are essential for ensuring the security and reliability of industrial controller software. Fuzzing, a technique that automatically discovers software bugs, has also proven effective in finding software vulnerabilities. Unsurprisingly, fuzzing has been applied to a wide range of platforms, including programmable logic controllers (PLCs). However, current approaches, such as coverage-guided evolutionary fuzzing implemented in the popular fuzzer American Fuzzy Lop Plus Plus (AFL++), are often inadequate for finding logical errors and bugs in PLC control logic applications. They primarily target generic programming languages like C/C++, Java, and Python, and do not consider the unique characteristics and behaviors of PLCs, which are often programmed using specialized programming languages like Structured Text (ST). Furthermore, these fuzzers are ill suited to deal with complex input structures encapsulated in ST, as they are not specifically designed to generate appropriate input sequences. This renders the application of traditional fuzzing techniques less efficient on these platforms. To address this issue, this paper presents a fuzzing framework designed explicitly for PLC software to discover logic bugs in applications written in ST specified by the IEC 61131-3 standard. The proposed framework incorporates a custom-tailored PLC runtime and a fuzzer designed for the purpose. We demonstrate its effectiveness by fuzzing a collection of ST programs that were crafted for evaluation purposes. We compare the performance against a popular fuzzer, namely, AFL++. The proposed fuzzing framework demonstrated its capabilities in our experiments, successfully detecting logic bugs in the tested PLC control logic applications written in ST. On average, it was at least 83 times faster than AFL++, and in certain cases, for example, it was more than 23,000 times faster.

Fuzzing is an automated process for discovering inputs in a program that may trigger unexpected behavior. Today, fuzzing has become a standard practice for the discovery of bugs and security vulnerabilities. However, the main issue with such practices is that the exploration of the input space of programs can often be prohibitively expensive. Therefore, several alternative fuzzing strategies have been introduced during the last few years. Some fuzzing techniques rely on human expertise to provide a plausible set of initial input examples, namely, seeds. However, the process of handcrafting seeds for fuzzing purposes often becomes strenuous for humans as it requires a deeper understanding of the Program-Under-Test (PUT). Also, the use of known inputs to programs often does not trigger vulnerable program behavior or may not reach potentially vulnerable code locations. To address those issues, we propose a seed generation framework that enables Human-In-The-Loop (HITL) directed fuzzing where the human assumes a more active role in the creation of seeds that can penetrate and assess desired locations of the PUT. Our proposed framework uses Symbolic Execution (SE) to generate seeds that exercise paths to target program locations. Moreover, our framework enables the visualization of the explored execution paths in the binary of the PUT for the generated seeds. We evaluated our approach on a set of 12 carefully designed C programs with diverse characteristics that mimic real-world programs. The experimental results show the effectiveness of the proposed approach in improving the performance of standard fuzzing tools such as the American Fuzzy Lop ("Image missing"). Specifically, our solution can generate seeds that substantially enhance the performance of the fuzzer, achieving speedups ranging from $$1.46\times $$ 1.46 × to $$68.53\times $$ 68.53 × for branch conditions, $$1.39\times $$ 1.39 × to $$254.62\times $$ 254.62 × for branch depths, $$14,879.59\times $$ 14 , 879.59 × to $$30,295.88\times $$ 30 , 295.88 × for branch widths over traditional seeds. Additionally, the speedup increases with the number of target function ranging from $$12,260\times $$ 12 , 260 × to $$22,856.07\times $$ 22 , 856.07 × over traditional seeds while only requiring less than 15 seconds on average for the seed generation step.

Network security tools are indispensable in testing and evaluating the security of computer networks. Existing tools, such as Hping3, however, offer a limited set of options and attack-specific configurations, which restrict their use solely to well-known attack patterns. Although highly parameterizable libraries, such as Scapy, provide more options and scripting capabilities, they require extensive manual setup and often a steep learning curve. The development of powerful AI models, capitalizing on the transformer architecture, has enabled cybersecurity researchers to develop or incorporate these models into existing cyber-defense systems and red-team assessments. Prominent models such as NetGPT, TrafficFormer, and TrafficGPT can be effective, but require extensive computational resources for fine-tuning and a complex setup to adapt to proprietary networking environments and protocols. In this work, we propose AgentRed, a lightweight tool for generating network attack traffic with minimal human configuration and setup. Our tool integrates an AI agent and a large language model with fewer than a billion parameters into the network traffic generation process. Our method creates lightweight Low-Rank Adaptation (LoRA) adapters that can learn specific traffic patterns in a particular network environment. Our agent can autonomously train the LoRA adapters, search online documentation for attack patterns and parameters, and select appropriate adapters to generate network traffic specific to the user’s needs. It utilizes the LoRA adapters to create an intermediate traffic representation that can be parsed and executed by tools such as Scapy to generate malicious traffic in a virtualized test environment. We assess the performance of the proposed approach on six popular network attacks, including flooding attacks, Smurf, Ping-of-Death, and normal ICMP ping traffic. Our results validate the ability of the proposed tool to efficiently generate network packets with 97.9% accuracy using the LoRA adapters, compared to 95.4% accuracy using the base pre-trained Qwen3 0.6B model. When the AI agent performs online searches to enrich the LoRA adapters’ context during traffic generation, our method maintains an accuracy of 96.0% across all tested traffic patterns.