Online scams continue to cause substantial financial and personal harm. As a result, detection systems based on Large Language Models (LLMs) have been integrated into security products ranging from email gateways and browser extensions to fraud-monitoring dashboards. As this adoption accelerates, a common belief has taken hold: that these models are broadly suitable for scam detection. In this work, we investigate whether LLMs, with their strong capabilities in understanding intent, context, and reasoning, can effectively detect scams across diverse scenarios without task-specific fine-tuning. We curate and release a unique benchmark dataset of real-world scams spanning multiple formats and topics. We evaluate nine LLMs of varying sizes and architectures, examining their performance under different prompting strategies and comparing them to a fine-tuned BERT-based classifier. Our results show that while larger LLMs generally outperform smaller ones, effective prompting substantially boosts the performance of smaller models. Moreover, LLMs are better at generalizing to unseen scams compared to fine-tuned models, suggesting that pre-trained knowledge contributes meaningfully to scam detection. We release our dataset and evaluation framework to facilitate future research in robust scam detection using language models.

The Better Business Bureau ranked online e-commerce scams as the top consumer threat for 2022. Our measurements of real consumer devices confirm that e-commerce scams receive large traffic volumes, a total of 6.3M visits during seven months. In this work, we study e-commerce scams in depth and design a detection classifier that combines novel features that target salient characteristics of e-commerce scam websites and features for detecting malicious and scam domains proposed by prior work. In addition, we specify a method for automatically creating reliable ground-truth sets that are an order of magnitude larger than that of prior work. We use this data set to evaluate the classifier and achieve a high 0.973 F1-score (Prec: 0.988, Rec: 0.959). In a best-effort comparison, we demonstrate that our classifier outperforms the F-1 score of the prior art by 11% and that our novel features offer an F1-score boost of 4.3% over the features used in the prior art. In addition, we deploy our classifier in a real-world setting, analyze over 760K e-shops visited by real users, and identify 10% of those as e-commerce scams. We demonstrate that the classifier has a low False Positive rate in real-world settings and can protect over 176K users in one week.

Currently, WHOIS is the main method for identifying which company or individual owns a domain or website. But, WHOIS usefulness is limited due to privacy protection services and data redaction. We present a novel automated approach for domain and website attribution. When WHOIS data does not reveal the owner, our approach leverages information from multiple other sources such as passive DNS, TLS certificates, and the analysis of website content. We propose a novel ranking technique to select the domain owner among multiple identified entities. Our approach identifies the domain owner with an F1 score of 0.94 compared to 0.54 for WHOIS. When applied on 3,001 tracker domains from the popular Disconnect list, it identifies needed updates to the list. It also attributes 84% of previously unattributed tracker domains.

Many studies have proposed machine-learning (ML) models for malware detection and classification, reporting an almost-perfect performance. However, they assemble ground-truth in different ways, use diverse static- and dynamic-analysis techniques for feature extraction, and even differ on what they consider a malware family. As a consequence, our community still lacks an understanding of malware classification results: whether they are tied to the nature and distribution of the collected dataset, to what extent the number of families and samples in the training dataset influence performance, and how well static and dynamic features complement each other.

Malware is still a widespread problem, and it is used by malicious actors to routinely compromise the security of computer systems. Consumers typically rely on a single AV product to detect and block possible malware infections, while corporations often install multiple security products, activate several layers of defenses, and establish security policies among employees. However, if a better security posture should lower the risk of malware infections, then the actual extent to which this happens is still under debate by risk analysis experts. Moreover, the difference in risks encountered by consumers and enterprises has never been empirically studied by using real-world data. In fact, the mere use of third-party software, network services, and the interconnected nature of our society necessarily exposes both classes of users to undiversifiable risks: Independently from how careful users are and how well they manage their cyber hygiene, a portion of that risk would simply exist because of the fact of using a computer, sharing the same networks, and running the same software. In this work, we shed light on both systemic (i.e., diversifiable and dependent on the security posture) and systematic (i.e., undiversifiable and independent of the cyber hygiene) risk classes. Leveraging the telemetry data of a popular security company, we compare, in the first part of our study, the effects that different security measures have on malware encounter risks in consumer and enterprise environments. In the second part, we conduct exploratory research on systematic risk, investigate the quality of nine different indicators we were able to extract from our telemetry, and provide, for the first time, quantitative indicators of their predictive power. Our results show that even if consumers have a slightly lower encounter rate than enterprises (9.8% vs. 12.0%), the latter do considerably better when selecting machines with an increasingly higher uptime (89% vs. 53%). The two segments also diverge when we separately consider the presence of Adware and Potentially Unwanted Applications (PUA) and the generic samples detected through behavioral signatures: While consumers have an encounter rate for Adware and PUA that is 6 times higher than enterprise machines, those on average match behavioral signatures 2 times more frequently than the counterpart. We find, instead, similar trends when analyzing the age of encountered signatures, and the prevalence of different classes of traditional malware (such as Ransomware and Cryptominers). Finally, our findings show that the amount of time a host is active, the volume of files generated on the machine, the number and reputation of vendors of the installed applications, the host geographical location, and its recurrent infected state carry useful information as indicators of systematic risk of malware encounters. Activity days and hours have a higher influence in the risk of consumers, increasing the odds of encountering malware of 4.51 and 2.65 times. In addition, we measure that the volume of files generated on the host represents a reliable indicator, especially when considering Adware. We further report that the likelihood of encountering Worms and Adware is much higher (on average 8 times in consumers and enterprises) for those machines that already reported this kind of signature in the past.

Web pages have been steadily increasing in complexity over time, including code snippets from several distinct origins and organizations. While this may be a known phenomenon, its implications on the panorama of cookie tracking received little attention until now. Our study focuses on filling this gap, through the analysis of crawl results that are both large-scale and fine-grained, encompassing the whole set of events that lead to the creation and sharing of around 138 million cookies from crawling more than 6 million webpages. Our analysis lets us paint a highly detailed picture of the cookie ecosystem, discovering an intricate network of connections between players that reciprocally exchange information and include each other's content in web pages whose owners may not even be aware. We discover that, in most webpages, tracking cookies are set and shared by organizations at the end of complex chains that involve several middlemen. We also study the impact of cookie ghostwriting, i.e., a common practice where an entity creates cookies in the name of another party, or the webpage. We attribute and define a set of roles in the cookie ecosystem, related to cookie creation and sharing. We see that organizations can and do follow different patterns, including behaviors that previous studies could not uncover: for example, many cookie ghostwriters send cookies they create to themselves, which makes them able to perform cross-site tracking even for users that deleted third-party cookies in their browsers. While some organizations concentrate the flow of information on themselves, others behave as dispatchers, allowing other organizations to perform tracking on the pages that include their content.

Android is the most popular operating system with billions of active devices. Unfortunately, its popularity and openness makes it attractive for unwanted apps, i.e., malware and potentially unwanted programs (PUP). In Android, app installations typically happen via the official and alternative markets, but also via other smaller and less understood alternative distribution vectors such as Web downloads, pay-per-install (PPI) services, backup restoration, bloatware, and IM tools. This work performs a thorough investigation on unwanted app distribution by quantifying and comparing distribution through different vectors. At the core of our measurements are reputation logs of a large security vendor, which include 7.9M apps observed in 12M devices between June and September 2019. As a first step, we measure that between 10% and 24% of users devices encounter at least one unwanted app, and compare the prevalence of malware and PUP. An analysis of the who-installs-who relationships between installers and child apps reveals that the Play market is the main app distribution vector, responsible for 87% of all installs and 67% of unwanted app installs, but it also has the best defenses against unwanted apps. Alternative markets distribute instead 5.7% of all apps, but over 10% of unwanted apps. Bloatware is also a significant unwanted app distribution vector with 6% of those installs. And, backup restoration is an unintentional distribution vector that may even allow unwanted apps to survive users’ phone replacement. We estimate unwanted app distribution via PPI to be smaller than on Windows. Finally, we observe that Web downloads are rare, but provide a riskier proposition even compared to alternative markets.

The recent emergence of consumer off-the-shelf embedded (IoT) devices and the rise of large-scale IoT botnets has dramatically increased the volume and sophistication of Linux malware observed in the wild. The security community has put a lot of effort to document these threats but analysts mostly rely on manual work, which makes it difficult to scale and hard to regularly maintain. Moreover, the vast amount of code reuse that characterizes IoT malware calls for an automated approach to detect similarities and identify the phylogenetic tree of each family.

Cyber attacks have increased in number and complexity in recent years, and companies and organizations have accordingly raised their investments in more robust infrastructure to preserve their data, assets and reputation. However, the full protection against these countless and constantly evolving threats is unattainable by the sole use of preventive measures. Therefore, to handle residual risks and contain business losses in case of an incident, firms are increasingly adopting a cyber insurance as part of their corporate risk management strategy.As a result, the cyber insurance sector - which offers to transfer the financial risks related to network and computer incidents to a third party - is rapidly growing, with recent claims that already reached a $100M dollars. However, while other insurance sectors rely on consolidated methodologies to accurately predict risks, the many peculiarities of the cyber domain resulted in carriers to often resort to qualitative approaches based on experts opinions.This paper looks at past research conducted in the area of cyber insurance and classifies previous studies in four different areas, focused respectively on studying the economical aspects, the mathematical models, the risk management methodologies, and the predictions of cyber events. We then identify, for each insurance phase, a group of practical research problems where security experts can help develop new data-driven methodologies and automated tools to replace the existing qualitative approaches.

The European Union's (EU) General Data Protection Regulation (GDPR), in effect since May 2018, enforces strict limitations on handling users' personal data, hence impacting their activity tracking on the Web. In this study, we perform an evaluation of the tracking performed in 2,000 high-traffic websites, hosted both inside and outside of the EU. We evaluate both the information presented to users and the actual tracking implemented through cookies; we find that the GDPR has impacted website behavior in a truly global way, both directly and indirectly: USA-based websites behave similarly to EU-based ones, while third-party opt-out services reduce the amount of tracking even for websites which do not put any effort in respecting the new law. On the other hand, we find that tracking remains ubiquitous. In particular, we found cookies that can identify users when visiting more than 90% of the websites in our dataset - and we also encountered a large number of websites that present deceiving information, making it it very difficult, if at all possible, for users to avoid being tracked.

Most websites, services, and applications have come to rely on Internet services (e.g., DNS, CDN, email, WWW, etc.) offered by third parties. Although employing such services generally improves reliability and cost-effectiveness, it also creates dependencies on service providers, which may expose websites to additional risks, such as DDoS attacks or cascading failures. As cloud services are becoming more popular, an increasing percentage of the overall Internet ecosystem relies on a decreasing number of highly popular services. In our general effort to assess the security risk for a given entity, and motivated by the effects of recent service disruptions, we perform a large-scale analysis of passive and active DNS datasets including more than 2.5 trillion queries in order to discover the dependencies between websites and Internet services.

The current evolution of the cyber-threat ecosystem shows that no system can be considered invulnerable. It is therefore important to quantify the risk level within a system and devise risk prediction methods such that proactive measures can be taken to reduce the damage of cyber attacks. We present RiskTeller, a system that analyzes binary file appearance logs of machines to predict which machines are at risk of infection months in advance. Risk prediction models are built by creating, for each machine, a comprehensive profile capturing its usage patterns, and then associating each profile to a risk level through both fully and semi-supervised learning methods. We evaluate RiskTeller on a year-long dataset containing information about all the binaries appearing on machines of 18 enterprises. We show that RiskTeller can use the machine profile computed for a given machine to predict subsequent infections with the highest prediction precision achieved to date.

Malware remains an important security threat, as miscreants continue to deliver a variety of malicious programs to hosts around the world. At the heart of all the malware delivery techniques are executable files (known as downloader trojans or droppers) that download other malware. Because the act of downloading software components from the Internet is not inherently malicious, benign and malicious downloaders are difficult to distinguish based only on their content and behavior. In this paper, we introduce the downloader-graph abstraction, which captures the download activity on end hosts, and we explore the growth patterns of benign and malicious graphs. Downloader graphs have the potential of exposing large parts of the malware download activity, which may otherwise remain undetected. By combining telemetry from anti-virus and intrusion-prevention systems, we reconstruct and analyze 19 million downloader graphs from 5 million real hosts. We identify several strong indicators of malicious activity, such as the growth rate, the diameter, and the Internet access patterns of downloader graphs. Building on these insights, we implement and evaluate a machine learning system for malware detection. Our system achieves a 96.0% true-positive rate, with a 1.0% false-positive rate, and detects malware an average of 9.24 days earlier than existing anti-virus products. We also perform an external validation by examining a sample of unlabeled files that our system detects as malicious, and we find that 41.41% are blocked by anti-virus products.

Vulnerability exploits remain an important mechanism for malware delivery, despite efforts to speed up the creation of patches and improvements in software updating mechanisms. Vulnerabilities in client applications (e.g., Browsers, multimedia players, document readers and editors) are often exploited in spear phishing attacks and are difficult to characterize using network vulnerability scanners. Analyzing their lifecycle requires observing the deployment of patches on hosts around the world. Using data collected over 5 years on 8.4 million hosts, available through Symantec's WINE platform, we present the first systematic study of patch deployment in client-side vulnerabilities. We analyze the patch deployment process of 1,593 vulnerabilities from 10 popular client applications, and we identify several new threats presented by multiple installations of the same program and by shared libraries distributed with several applications. For the 80 vulnerabilities in our dataset that affect code shared by two applications, the time between patch releases in the different applications is up to 118 days (with a median of 11 days). Furthermore, as the patching rates differ considerably among applications, many hosts patch the vulnerability in one application but not in the other one. We demonstrate two novel attacks that enable exploitation by invoking old versions of applications that are used infrequently, but remain installed. We also find that the median fraction of vulnerable hosts patched when exploits are released is at most 14%. Finally, we show that the patching rate is affected by user-specific and application-specific factors, for example, hosts belonging to security analysts and applications with an automated updating mechanism have significantly lower median times to patch.

Users are typically the final target of web attacks: criminals are interested in stealing their money, their personal information, or in infecting their machines with malicious code. However, while many aspects of web attacks have been carefully studied by researchers and security companies, the reasons that make certain users more "at risk" than others are still unknown. Why do certain users never encounter malicious pages while others seem to end up on them on a daily basis?

A wide range of malicious activities rely on the domain name service (DNS) to manage their large, distributed networks of infected machines. As a consequence, the monitoring and analysis of DNS queries has recently been proposed as one of the most promising techniques to detect and blacklist domains involved in malicious activities (e.g., phishing, spam, botnets command-and-control, etc.). EXPOSURE is a system we designed to detect such domains in real time, by applying 15 unique features grouped in four categories. We conducted a controlled experiment with a large, real-world dataset consisting of billions of DNS requests. The extremely positive results obtained in the tests convinced us to implement our techniques and deploy it as a free, online service. In this article, we present the Exposure system and describe the results and lessons learned from 17 months of its operation. Over this amount of time, the service detected over 100K malicious domains. The statistics about the time of usage, number of queries, and target IP addresses of each domain are also published on a daily basis on the service Web page.

Botnets continue to be a significant problem on the Internet. Accordingly, a great deal of research has focused on methods for detecting and mitigating the effects of botnets. Two of the primary factors preventing the development of effective large-scale, wide-area botnet detection systems are seemingly contradictory. On the one hand, technical and administrative restrictions result in a general unavailability of raw network data that would facilitate botnet detection on a large scale. On the other hand, were this data available, real-time processing at that scale would be a formidable challenge. In contrast to raw network data, NetFlow data is widely available. However, NetFlow data imposes several challenges for performing accurate botnet detection.

Little is known about the duration and prevalence of zero-day attacks, which exploit vulnerabilities that have not been disclosed publicly. Knowledge of new vulnerabilities gives cyber criminals a free pass to attack any target of their choosing, while remaining undetected. Unfortunately, these serious threats are difficult to analyze, because, in general, data is not available until after an attack is discovered. Moreover, zero-day attacks are rare events that are unlikely to be observed in honeypots or in lab experiments.

Despite the numerous prevention and protection mechanisms that have been introduced into modern operating systems, the exploitation of memory corruption vulnerabilities still represents a serious threat to the security of software systems and networks. A recent exploitation technique, called Return-Oriented Programming (ROP), has lately attracted a considerable attention from academia. Past research on the topic has mostly focused on refining the original attack technique, or on proposing partial solutions that target only particular variants of the attack.

Social networking sites have been increasingly gaining popularity. Well-known sites such as Facebook have been reporting growth rates as high as 3% per week. Many social networking sites have millions of registered users who use these sites to share photographs, contact long-lost friends, establish new business contacts and to keep in touch. In this paper, we investigate how easy it would be for a potential attacker to launch automated crawling and identity theft attacks against a number of popular social networking sites in order to gain access to a large volume of personal user information. The first attack we present is the automated identity theft of existing user profiles and sending of friend requests to the contacts of the cloned victim. The hope, from the attacker's point of view, is that the contacted users simply trust and accept the friend request. By establishing a friendship relationship with the contacts of a victim, the attacker is able to access the sensitive personal information provided by them. In the second, more advanced attack we present, we show that it is effective and feasible to launch an automated, cross-site profile cloning attack. In this attack, we are able to automatically create a forged profile in a network where the victim is not registered yet and contact the victim's friends who are registered on both networks. Our experimental results with real users show that the automated attacks we present are effective and feasible in practice.