Alexandros Kapravelos

GitHub, the de facto platform for open-source software development, provides a set of social-media-like features to signal high-quality repositories. Among them, the star count is the most widely used popularity signal, but it is also at risk of being artificially inflated (i.e., faked), decreasing its value as a decision-making signal and posing a security risk to all GitHub users. In this paper, we present a systematic, global, and longitudinal measurement study of fake stars in GitHub. To this end, we build StarScout, a scalable tool able to detect anomalous starring behaviors across all GitHub metadata between 2019 and 2024. Analyzing the data collected using StarScout, we find that: (1) fake-star-related activities have rapidly surged in 2024; (2) the accounts and repositories in fake star campaigns have highly trivial activity patterns; (3) the majority of fake stars are used to promote short-lived phishing malware repositories; the remaining ones are mostly used to promote AI/LLM, blockchain, tool/application, and tutorial/demo repositories; (4) while repositories may have acquired fake stars for growth hacking, fake stars only have a promotion effect in the short term (i.e., less than two months) and become a liability in the long term. Our study has implications for platform moderators, open-source practitioners, and supply chain security researchers.

Large language model (LLM)-based AI agents extend LLM capabilities by enabling access to tools such as data sources, APIs, search engines, code sandboxes, and even other agents. While this empowers agents to perform complex tasks, LLMs may invoke unintended tool interactions and introduce risks, such as leaking sensitive data or overwriting critical records, which are unacceptable in enterprise contexts. Current approaches to mitigate these risks, such as model-based safeguards, enhance agents’ reliability but cannot guarantee system safety. Methods like information flow control (IFC) and temporal constraints aim to provide guarantees but often require extensive human annotation. We propose a process that starts with applying System-Theoretic Process Analysis (STPA) to identify hazards in agent workflows, derive safety requirements, and formalize them as enforceable specifications on data flows and tool sequences. To enable this, we introduce a capability-enhanced Model Context Protocol (MCP) framework that requires structured labels on capabilities, confidentiality, and trust level. Together, these contributions aim to shift LLM-based agent safety from ad hoc reliability fixes to proactive guardrails with formal guarantees, while reducing dependence on user confirmation and making autonomy a deliberate design choice.

Static Application Security Testing (SAST) tools using taint analysis are widely viewed as providing higher-quality vulnerability detection results compared to traditional pattern-based approaches. However, performing static taint analysis for JavaScript poses two major challenges. First, JavaScript's dynamic features complicate data flow extraction required for taint tracking. Second, npm's large library ecosystem makes it difficult to identify relevant sources/sinks and establish taint propagation across dependencies. In this paper, we present SemTaint, a multi-agent system that strategically combines the semantic understanding of Large Language Models (LLMs) with traditional static program analysis to extract taint specifications, including sources, sinks, call edges, and library flow summaries tailored to each package. Conceptually, SemTaint uses static program analysis to calculate a call graph and defers to an LLM to resolve call edges that cannot be resolved statically. Further, it uses the LLM to classify sources and sinks for a given CWE. The resulting taint specification is then provided to a SAST tool, which performs vulnerability analysis. We integrate SemTaint with CodeQL, a state-of-the-art SAST tool, and demonstrate its effectiveness by detecting 106 of 162 vulnerabilities previously undetectable by CodeQL. Furthermore, we find 4 novel vulnerabilities in 4 popular npm packages. In doing so, we demonstrate that LLMs can practically enhance existing static program analysis algorithms, combining the strengths of both symbolic reasoning and semantic understanding for improved vulnerability detection.

In 2024, the Anti-Phishing Work Group identified over one million phishing pages. Phishers achieve this scale by using phishing kits -- ready-to-deploy phishing websites -- to rapidly deploy phishing campaigns with specific data exfiltration, evasion, or mimicry techniques. In contrast, researchers and defenders continue to fight phishing on a page-by-page basis and rely on manual analysis to recognize static features for kit identification. This paper aims to aid researchers and analysts by automatically differentiating groups of phishing pages based on the underlying kit, automating a previously manual process, and enabling us to measure how popular different client-side techniques are across these groups. For kit detection, our system has an accuracy of 97% on a ground-truth dataset of 548 kit families deployed across 4,562 phishing URLs. On an unlabeled dataset, we leverage the complexity of 434,050 phishing pages' JavaScript logic to group them into 11,377 clusters, annotating the clusters with what phishing techniques they employ. We find that UI interactivity and basic fingerprinting are universal techniques, present in 90% and 80% of the clusters, respectively. On the other hand, mouse detection via the browser's mouse API is among the rarest behaviors, despite being used in a deployment of a 7-year-old open-source phishing kit. Our methods and findings provide new ways for researchers and analysts to tackle the volume of phishing pages.

Reusable software libraries, frameworks, and components, such as those provided by open source ecosystems and third-party suppliers, accelerate digital innovation. However, recent years have shown almost exponential growth in attackers leveraging these software artifacts to launch software supply chain attacks. Past well-known software supply chain attacks include the SolarWinds, log4j, and xz utils incidents. Supply chain attacks are considered to have three major attack vectors: through vulnerabilities and malware accidentally or intentionally injected into open source and third-party dependencies/components/containers ; by infiltrating the build infrastructure during the build and deployment processes; and through targeted techniques aimed at the humans involved in software development, such as through social engineering. Plummeting trust in the software supply chain could decelerate digital innovation if the software industry reduces its use of open source and third-party artifacts to reduce risks. This article contains perspectives and knowledge obtained from intentional outreach with practitioners to understand their practical challenges and from extensive research efforts. We then provide an overview of current research efforts to secure the software supply chain. Finally, we propose a future research agenda to close software supply chain attack vectors and support the software industry.

The powerful capabilities of modern browsers have pushed the web application logic to the user side, in order to minimize latency, increase scalability of the service and improve users' quality of experience. What is more, browsers provide a rich toolchest for browser extensions to provide additional functionality, but at the same time enable them to become a powerful vehicle for malicious actors. Such actors may spy, phish or fraud users, thus making the user's browser untrusted for the web servers. In this paper, we present WRIT, a practical framework that enables websites to protect critical functionality from abuse in the presence of malicious extensions. In WRIT, the integrity of outgoing web requests is attested and verified to ensure they were triggered by a user's action and not automatically generated by a malicious browser extension. WRIT is immediately applicable by leveraging existing HTML5 and other native browser features and does not require any modification of the browser. Performance results of our prototype show that it adds a negligible 7.29 ms latency to sensitive user-triggered actions (e.g., post message).

With the progression of modern browsers, online tracking has become the most concerning issue for preserving privacy on the web. As major browser vendors plan to or already ban third-party cookies, trackers have to shift towards browser fingerprinting by incorporating novel browser APIs into their tracking arsenal. Understanding how new browser APIs are abused in browser fingerprinting techniques is a significant step toward ensuring protection from online tracking. In this paper, we propose a novel hybrid system, named BFAD, that automatically identifies previously unknown browser fingerprinting APIs in the wild. The system combines dynamic and static analysis to accurately reveal browser API usage and automatically infer browser fingerprinting behavior. Based on the observation that a browser fingerprint is constructed by pulling information from multiple APIs, we leverage dynamic analysis and a locality-based algorithm to discover all involved APIs and static analysis on the dataflow of fingerprinting information to accurately associate them together. Our system discovers 231 fingerprinting APIs in Alexa top 10K domains, starting with only 35 commonly known fingerprinting APIs and 17 data transmission APIs. Out of 231 APIs, 161 of them are not identified by state-of-the-art detection systems. Since our approach is fully automated, we repeat our experiments 11 months later and discover 18 new fingerprinting APIs that were not discovered in our previous experiment. We present with case studies the fingerprinting ability of a total of 249 detected APIs.