Penetration testing, a critical component of cybersecurity, typically requires extensive time and effort to find vulnerabilities. Beginners in this field often benefit from collaborative approaches with the community or experts. To address this, we develop Cybersecurity Intelligent Penetration-testing Helper for Ethical Researchers (CIPHER), a large language model specifically trained to assist in penetration testing tasks as a chatbot. Unlike software development, penetration testing involves domain-specific knowledge that is not widely documented or easily accessible, necessitating a specialized training approach for AI language models. CIPHER was trained using over 300 high-quality write-ups of vulnerable machines, hacking techniques, and documentation of open-source penetration testing tools augmented in an expert response structure. Additionally, we introduced the Findings, Action, Reasoning, and Results (FARR) Flow augmentation, a novel method to augment penetration testing write-ups to establish a fully automated pentesting simulation benchmark tailored for large language models. This approach fills a significant gap in traditional cybersecurity Q&A benchmarks and provides a realistic and rigorous standard for evaluating LLM’s technical knowledge, reasoning capabilities, and practical utility in dynamic penetration testing scenarios. In our assessments, CIPHER achieved the best overall performance in providing accurate suggestion responses compared to other open-source penetration testing models of similar size and even larger state-of-the-art models like Llama 3 70B and Qwen1.5 72B Chat, particularly on insane difficulty machine setups. This demonstrates that the current capabilities of general large language models (LLMs) are insufficient for effectively guiding users through the penetration testing process. We also discuss the potential for improvement through scaling and the development of better benchmarks using FARR Flow augmentation results.

Detecting anomalies in industrial sound is critical for maintaining operational efficiency, preventing costly equipment failures, and ensuring workplace safety. However, it presents significant challenges due to the complexity and variability of industrial environments, including background noise and fluctuating operating conditions. This paper proposes a comprehensive approach that leverages machine learning (ML) and deep learning (DL) techniques to address these challenges. Using three datasets, the Malfunctioning Industrial Machine Investigation and Inspection for Domain Generalization 2022 (MIMII DG 2022), the Detection and Classification of Acoustic Scenes and Events (DCASE) 2022, and DCASE 2024, we evaluate the performance of various ML and DL models under different experimental conditions. Our study focuses on feature extraction methods such as Mel-spectrograms and Mel-frequency Cepstral Coefficients (MFCCs) to capture critical acoustic characteristics of industrial machinery. Both supervised ML and DL techniques are employed to explore effective anomaly detection strategies. Extensive experimentation and evaluation using metrics such as confusion matrix, accuracy, precision, recall, and F1 score highlight the effectiveness of our approach in real-world industrial scenarios. Experimental results demonstrate that eXtreme Gradient Boosting (XGBoost) outperforms Support Vector Machine (SVM) and Decision Tree (DT) models in the ML approach across both feature sets. In the DL approach, Gated Recurrent Units (GRU) perform better on MFCC features than Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM) networks. GRU emerges as the best-performing model on testing datasets, achieving a precision of 99.56% and an F1 score of 99.55% on the DCASE 2024 dataset, a precision of 94% and an F1 score of 93.95% on the DCASE 2022 dataset, and a precision of 92.2% and an F1 score of 92.06% on the MIMII DG 2022 dataset. These results underscore the potential of DL for real-time industrial sound analysis in predictive maintenance, offering significant improvements over traditional methods.

The Controller Area Network (CAN) bus plays a key role in keeping vehicles safe by enabling critical systems to communicate with each other. However, because it does not have its own security features, the CAN bus is open to cyber threats. A CAN bus intrusion detection system (IDS) is critical for automotive cybersecurity. This has made it especially important to create IDS that are not just accurate but also efficient enough to run on the limited hardware of Electronic Control Units (ECUs). Unfortunately, many current deep learning solutions for CAN intrusion detection use large and complex models that are too demanding for most automotive systems. Moreover, existing deep learning approaches need excessive computational resources that are unsuitable for resource-constrained ECUs. We propose TinyCNNCANNet, an ultra-lightweight convolutional neural network with just 13K parameters, designed to provide low-latency and resource-efficient CAN intrusion detection under experimental settings. Rather than focusing on on-vehicle deployment, this work evaluates the feasibility of lightweight CNN architectures for future real-time capable CAN intrusion detection. We comprehensively evaluate TinyCNNCANNet on four diverse datasets: CANFD 2021, CICIoV 2024, Multi-Fuzzer-CAN 2025, and SynCAN 2025. These datasets encompass nine attack types. TinyCNNCANNet achieves competitive or superior performance compared to models with 115- $300\times $ more parameters. All architectures detect volume-based attacks (DoS, flooding, and fuzzing) most effectively. Sophisticated attacks (malfunction and fuzzer variants) challenge all models to a similar degree, regardless of complexity. TinyCNNCANNet shows superior generalization on synthetic out-of-distribution data (SynCAN 2025). It achieves 100% accuracy, while EfficientCANNet (86.82%) and MobileNetCANNet (59.33%) fail, revealing overfitting vulnerabilities in complex models. TinyCNNCANNet delivers 12- $20\times $ faster inference (0.16-0.51 ms vs. 2.14-4.15 ms) and a 145- $383\times $ smaller model size (0.04 MB vs. 5.81-15.32 MB). These results demonstrate the potential of TinyCNNCANNet for real-time capable CAN intrusion detection and indicate its suitability for future deployment on embedded automotive platforms.

The rapid proliferation of unmanned aerial vehicles (UAVs) has created an urgent need for robust detection & identification systems to ensure airspace security and safety. However, the scarcity of large-scale, diverse, and accurately annotated real-world datasets hinders the development of effective UAV detection & identification algorithms. To address this challenge, we present the RV-DroneEye (Realistic Virtual DroneEye) Dataset, a comprehensive framework for generating synthetic datasets utilizing the Unity 3D simulation engine to enhance UAV detection training. Additionally, we leverage photorealistic augmentation using Flux.1 diffusion-based model as a base, diverse environmental conditions, and physically accurate flight dynamics to generate large-scale annotated datasets that capture the visual complexity of real-world UAV detection scenarios. Then we evaluate our dataset’s generalization on various benchmarks and train multiple object-detection models. The RV-DroneEye dataset includes diverse UAV models, environments (urban, forest, lake), weather, and lighting conditions.We assess its effectiveness by training state-of-the-art object detection models and evaluating them on real-world test sets, including a multiclass task for more than 20 UAV types. Results show that models trained on RV-DroneEye achieve comparable or better UAV pattern generalization than those using limited real-world data, with marked gains in challenging scenarios such as low-light conditions, extreme angles, and cluttered backgrounds.

Monocular depth estimation (MDE) and semantic segmentation (SS) are crucial for the navigation and environmental interpretation of many autonomous driving systems. However, their vulnerability to practical adversarial attacks is a significant concern. This paper presents a novel adversarial attack using practical patches that mimic manhole covers to deceive MDE and SS models. The goal is to cause these systems to misinterpret scenes, leading to false detections of near obstacles or non-passable objects. We use Depth Planar Mapping to precisely position these patches on road surfaces, enhancing the attack's effectiveness. Our experiments show that these adversarial patches cause a 43% relative error in MDE and achieve a 96% attack success rate in SS. These patches create affected error regions over twice their size in MDE and approximately equal to their size in SS. Our studies also confirm the patch's effectiveness in physical simulations, the adaptability of the patches across different target models, and the effectiveness of our proposed modules, highlighting their practical implications.