Maintaining the confidentiality, availability, and integrity of CI/CD systems against attacks which target their pipelines poses a significant security challenge. Attacks like Solarwinds's Orion Sunburst and CodeCov's Secret Exfiltration demonstrate that the abuse of simple misconfigurations spanning the vastly complex stack of services which make up CI/CD pipelines can have devastating consequences. In this paper, we quantify and consolidate CI/CD terminology, present a system architecture characterizing CI/CD, analyze patterns in existing CI/CD attacks, categorize popular free and open source defense tools, and perform gap analysis evaluation on existing tooling. We find that state-of-the-art CI/CD defense tools fail to mitigate 5 out of the 7 evaluation attacks we created.
Amid growing concerns over data privacy in web and mobile applications, this study aims to assess the privacy mechanisms in Instant Games on Facebook - a platform with approximately 3.03 billion monthly active users and a substantial repository of personal data. Instant Games have become increasingly popular due to their ease of access and social integration features. Investigating these games can provide insights into privacy mechanisms and practices, thereby informing the development of more fair, compliant and user privacy-centric gaming experiences. Thus, this paper proposes an integrated analytical framework that leverages a combination of descriptive, memory, and network analysis techniques to examine the privacy mechanisms in Facebook Instant Games. It focuses on evaluating the permission model, default settings, configurations, and API usage, as well as their impact on user data access, transfer, and sharing. Our findings uncovers discrepancies between privacy policies and actual user data notices. Through generalized settings and lack of explicit consent mechanism, our study reveals a system that often favors functionality over user privacy. Moreover, we highlight the reliance on powerful APIs that, while enhancing gameplay, pose additional privacy risks by granting broad data access to third-party services without direct user approval.
As threat modeling involves the architecture-centric identification, analysis and mitigation of security threats, it is considered an essential activity of the secure development lifecycle (SDLC). Threat modeling outcomes are used to guide the mitigation and risk management process, and to determine areas of focus in later development stages (e.g., testing and verification). Traditional threat modeling methods and techniques are driven by human experts and stakeholders and mainly conducted manually. Recently, focus has shifted towards automating (parts of) the process through improved tool support. In this paper, we evaluate the extent to which this trend also leads to easier development pipeline integration of threat modeling tools, i.e. whether this also enables conducting threat model-centric analysis within continuous integration and development (CI/CD) build pipelines so that they can contribute to the overall development-centric quality assurance process (SecDevOps). We first articulate the key requirements for threat modeling tools to be compatible or suited for CI/CD, and then we systematically evaluate and compare seven automated threat modeling tools against these requirements. Finally, we provide recommendations towards making threat analysis tools and enablers both more suitable for integration in a CI/CD context and hence, more compatible to modern software engineering processes.
Object detection stands as a fundamental component in numerous real-world applications, ranging from autonomous vehicles to security systems. As these technologies become increasingly embedded in our daily lives, ensuring the security and resilience of object detection models becomes critically essential. However, these models are vulnerable to adversarial attacks, where subtle alterations intentionally introduced into input data can mislead the model’s predictions. This paper explores the susceptibility of YOLO V8 and TensorFlow Object Detection models, such as MobileNet and ResNet, to adversarial attacks using the concept of attack transferability. Utilizing the Fast Gradient Sign Method alongside a distinct classifier model, we generate adversarial examples and evaluate the impact on object detection systems. Our analysis exposes a significant decline in object detection model’s performance in the presence of these adversarial examples, illustrating the transferability of attacks across different models. Our findings emphasize the critical necessity for robust defenses to safeguard object detection systems against transferable adversarial attacks.
Infrastructure-as-Code (IaC) is the practice of provisioning and managing cloud resources using machine-readable code. IaC is seeing increased adoption because it enhances transparency and reliability of infrastructure operations. However, as any software code, IaC can also contain misconfigurations, which can lead to insecure infrastructure, which may result in data breaches. Existing IaC scanning tools are able to detect common misconfigurations in IaC but they require IaC developers to manually repair the code. Recent advances in Large Language Models (LLMs) have led to promising results in applying LLMs to Automatic Program Repair (APR) tasks for code written in different languages. In this work, we propose an LLM-based approach to fix misconfigurations in IaC code. After misconfigurations in IaC code are identified by scanning tools, we feed the LLMs with the IaC code, details about the misconfigurations, and additional context provided by a human-in-the-loop and prompt the LLM to generate the repaired IaC code. We tested our approach on several vulnerable IaC repositories and found that the GPT-4 model from OpenAI suggests fixes that reduce up to 84.7% of the misconfiguration alarms produced by the scanners and our two-pass solution significantly improves the performance over a one-pass only approach. However, of the fixes suggested, we manually determined that only 79.6% actually solve the problem, while the remaining 20.4% are hallucinated fixes. Specifically, LLM hallucinations in the generated outputs pass checks for misconfigurations but fail other syntax and schema validation checks or do not address the underlying security issue. We propose a few potential approaches to tackle this challenge.
