William Robertson

Evan Li, Tushin Mallick, Evan Rose et al.

LLM-integrated app systems extend the utility of Large Language Models (LLMs) with third-party apps that are invoked by a system LLM using interleaved planning and execution phases to answer user queries. These systems introduce new attack vectors where malicious apps can cause integrity violation of planning or execution, availability breakdown, or privacy compromise during execution. In this work, we identify new attacks impacting the integrity of planning, as well as the integrity and availability of execution in LLM-integrated apps, and demonstrate them against IsolateGPT, a recent solution designed to mitigate attacks from malicious apps. We propose Abstract-Concrete-Execute (ACE), a new secure architecture for LLM-integrated app systems that provides security guarantees for system planning and execution. Specifically, ACE decouples planning into two phases by first creating an abstract execution plan using only trusted information, and then mapping the abstract plan to a concrete plan using installed system apps. We verify that the plans generated by our system satisfy user-specified secure information flow constraints via static analysis on the structured plan output. During execution, ACE enforces data and capability barriers between apps, and ensures that the execution is conducted according to the trusted abstract plan. We show experimentally that ACE is secure against attacks from the InjecAgent and Agent Security Bench benchmarks for indirect prompt injection, and our newly introduced attacks. We also evaluate the utility of ACE in realistic environments, using the Tool Usage suite from the LangChain benchmark. Our architecture represents a significant advancement towards hardening LLM-based systems using system security principles.

Georgios Syros, Evan Rose, Brian Grinstead et al.

Large language model (LLM) based web agents are increasingly deployed to automate complex online tasks by directly interacting with web sites and performing actions on users' behalf. While these agents offer powerful capabilities, their design exposes them to indirect prompt injection attacks embedded in untrusted web content, enabling adversaries to hijack agent behavior and violate user intent. Despite growing awareness of this threat, existing evaluations rely on fixed attack templates, manually selected injection surfaces, or narrowly scoped scenarios, limiting their ability to capture realistic, adaptive attacks encountered in practice. We present MUZZLE, an automated agentic framework for evaluating the security of web agents against indirect prompt injection attacks. MUZZLE utilizes the agent's trajectories to automatically identify high-salience injection surfaces, and adaptively generate context-aware malicious instructions that target violations of confidentiality, integrity, and availability. Unlike prior approaches, MUZZLE adapts its attack strategy based on the agent's observed execution trajectory and iteratively refines attacks using feedback from failed executions. We evaluate MUZZLE across diverse web applications, user tasks, and agent configurations, demonstrating its ability to automatically and adaptively assess the security of web agents with minimal human intervention. Our results show that MUZZLE effectively discovers 37 new attacks on 4 web applications with 10 adversarial objectives that violate confidentiality, availability, or privacy properties. MUZZLE also identifies novel attack strategies, including 2 cross-application prompt injection attacks and an agent-tailored phishing scenario.

Ioannis Agadakos, Manuel Egele, William Robertson

Designing and implementing secure software is inarguably more important than ever. However, despite years of research into privilege separating programs, it remains difficult to actually do so and such efforts can take years of labor-intensive engineering to reach fruition. At the same time, new intra-process isolation primitives make strong data isolation and privilege separation more attractive from a performance perspective. Yet, substituting intra-process security boundaries for time-tested process boundaries opens the door to subtle but devastating privilege leaks. In this work, we present Polytope, a language extension to C++ that aims to make efficient privilege separation accessible to a wider audience of developers. Polytope defines a policy language encoded as C++11 attributes that separate code and data into distinct program partitions. A modified Clang front-end embeds source-level policy as metadata nodes in the LLVM IR. An LLVM pass interprets embedded policy and instruments an IR with code to enforce the source-level policy using Intel MPK. A run-time support library manages partitions, protection keys, dynamic memory operations, and indirect call target privileges. An evaluation demonstrates that Polytope provides equivalent protection to prior systems with a low annotation burden and comparable performance overhead. Polytope also renders privilege leaks that contradict intended policy impossible to express.

Andrea Mambretti, Alexandra Sandulescu, Alessandro Sorniotti et al.

The prevalence of memory corruption bugs in the past decades resulted in numerous defenses, such as stack canaries, control flow integrity (CFI), and memory safe languages. These defenses can prevent entire classes of vulnerabilities, and help increase the security posture of a program. In this paper, we show that memory corruption defenses can be bypassed using speculative execution attacks. We study the cases of stack protectors, CFI, and bounds checks in Go, demonstrating under which conditions they can be bypassed by a form of speculative control flow hijack, relying on speculative or architectural overwrites of control flow data. Information is leaked by redirecting the speculative control flow of the victim to a gadget accessing secret data and acting as a side channel send. We also demonstrate, for the first time, that this can be achieved by stitching together multiple gadgets, in a speculative return-oriented programming attack. We discuss and implement software mitigations, showing moderate performance impact.

William Blair, Andrea Mambretti, Sajjad Arshad et al.

Contemporary fuzz testing techniques focus on identifying memory corruption vulnerabilities that allow adversaries to achieve either remote code execution or information disclosure. Meanwhile, Algorithmic Complexity (AC)vulnerabilities, which are a common attack vector for denial-of-service attacks, remain an understudied threat. In this paper, we present HotFuzz, a framework for automatically discovering AC vulnerabilities in Java libraries. HotFuzz uses micro-fuzzing, a genetic algorithm that evolves arbitrary Java objects in order to trigger the worst-case performance for a method under test. We define Small Recursive Instantiation (SRI) as a technique to derive seed inputs represented as Java objects to micro-fuzzing. After micro-fuzzing, HotFuzz synthesizes test cases that triggered AC vulnerabilities into Java programs and monitors their execution in order to reproduce vulnerabilities outside the fuzzing framework. HotFuzz outputs those programs that exhibit high CPU utilization as witnesses for AC vulnerabilities in a Java library. We evaluate HotFuzz over the Java Runtime Environment (JRE), the 100 most popular Java libraries on Maven, and challenges contained in the DARPA Space and Time Analysis for Cybersecurity (STAC) program. We evaluate SRI's effectiveness by comparing the performance of micro-fuzzing with SRI, measured by the number of AC vulnerabilities detected, to simply using empty values as seed inputs. In this evaluation, we verified known AC vulnerabilities, discovered previously unknown AC vulnerabilities that we responsibly reported to vendors, and received confirmation from both IBM and Oracle. Our results demonstrate that micro-fuzzing finds AC vulnerabilities in real-world software, and that micro-fuzzing with SRI-derived seed inputs outperforms using empty values.

Seyed Ali Mirheidari, Sajjad Arshad, Kaan Onarlioglu et al.

Web cache deception (WCD) is an attack proposed in 2017, where an attacker tricks a caching proxy into erroneously storing private information transmitted over the Internet and subsequently gains unauthorized access to that cached data. Due to the widespread use of web caches and, in particular, the use of massive networks of caching proxies deployed by content distribution network (CDN) providers as a critical component of the Internet, WCD puts a substantial population of Internet users at risk. We present the first large-scale study that quantifies the prevalence of WCD in 340 high-profile sites among the Alexa Top 5K. Our analysis reveals WCD vulnerabilities that leak private user data as well as secret authentication and authorization tokens that can be leveraged by an attacker to mount damaging web application attacks. Furthermore, we explore WCD in a scientific framework as an instance of the path confusion class of attacks, and demonstrate that variations on the path confusion technique used make it possible to exploit sites that are otherwise not impacted by the original attack. Our findings show that many popular sites remain vulnerable two years after the public disclosure of WCD. Our empirical experiments with popular CDN providers underline the fact that web caches are not plug & play technologies. In order to mitigate WCD, site operators must adopt a holistic view of their web infrastructure and carefully configure cache settings appropriate for their applications.

Sajjad Arshad, Amin Kharraz, William Robertson

Modern websites include various types of third-party content such as JavaScript, images, stylesheets, and Flash objects in order to create interactive user interfaces. In addition to explicit inclusion of third-party content by website publishers, ISPs and browser extensions are hijacking web browsing sessions with increasing frequency to inject third-party content (e.g., ads). However, third-party content can also introduce security risks to users of these websites, unbeknownst to both website operators and users. Because of the often highly dynamic nature of these inclusions as well as the use of advanced cloaking techniques in contemporary malware, it is exceedingly difficult to preemptively recognize and block inclusions of malicious third-party content before it has the chance to attack the user's system. In this paper, we propose a novel approach to achieving the goal of preemptive blocking of malicious third-party content inclusion through an analysis of inclusion sequences on the Web. We implemented our approach, called Excision, as a set of modifications to the Chromium browser that protects users from malicious inclusions while web pages load. Our analysis suggests that by adopting our in-browser approach, users can avoid a significant portion of malicious third-party content on the Web. Our evaluation shows that Excision effectively identifies malicious content while introducing a low false positive rate. Our experiments also demonstrate that our approach does not negatively impact a user's browsing experience when browsing popular websites drawn from the Alexa Top 500.

Muhammad Ahmad Bashir, Sajjad Arshad, William Robertson et al.

Numerous surveys have shown that Web users are concerned about the loss of privacy associated with online tracking. Alarmingly, these surveys also reveal that people are also unaware of the amount of data sharing that occurs between ad exchanges, and thus underestimate the privacy risks associated with online tracking. In reality, the modern ad ecosystem is fueled by a flow of user data between trackers and ad exchanges. Although recent work has shown that ad exchanges routinely perform cookie matching with other exchanges, these studies are based on brittle heuristics that cannot detect all forms of information sharing, especially under adversarial conditions. In this study, we develop a methodology that is able to detect client- and server-side flows of information between arbitrary ad exchanges. Our key insight is to leverage retargeted ads as a tool for identifying information flows. Intuitively, our methodology works because it relies on the semantics of how exchanges serve ads, rather than focusing on specific cookie matching mechanisms. Using crawled data on 35,448 ad impressions, we show that our methodology can successfully categorize four different kinds of information sharing behavior between ad exchanges, including cases where existing heuristic methods fail. We conclude with a discussion of how our findings and methodologies can be leveraged to give users more control over what kind of ads they see and how their information is shared between ad exchanges.

Sajjad Arshad, Amin Kharraz, William Robertson

Extensions provide useful additional functionality for web browsers, but are also an increasingly popular vector for attacks. Due to the high degree of privilege extensions can hold, extensions have been abused to inject advertisements into web pages that divert revenue from content publishers and potentially expose users to malware. Users are often unaware of such practices, believing the modifications to the page originate from publishers. Additionally, automated identification of unwanted third-party modifications is fundamentally difficult, as users are the ultimate arbiters of whether content is undesired in the absence of outright malice. To resolve this dilemma, we present a fine-grained approach to tracking the provenance of web content at the level of individual DOM elements. In conjunction with visual indicators, provenance information can be used to reliably determine the source of content modifications, distinguishing publisher content from content that originates from third parties such as extensions. We describe a prototype implementation of the approach called OriginTracer for Chromium, and evaluate its effectiveness, usability, and performance overhead through a user study and automated experiments. The results demonstrate a statistically significant improvement in the ability of users to identify unwanted third-party content such as injected ads with modest performance overhead.

Tobias Lauinger, Abdelberi Chaabane, Sajjad Arshad et al.

Web developers routinely rely on third-party Java-Script libraries such as jQuery to enhance the functionality of their sites. However, if not properly maintained, such dependencies can create attack vectors allowing a site to be compromised. In this paper, we conduct the first comprehensive study of client-side JavaScript library usage and the resulting security implications across the Web. Using data from over 133 k websites, we show that 37% of them include at least one library with a known vulnerability; the time lag behind the newest release of a library is measured in the order of years. In order to better understand why websites use so many vulnerable or outdated libraries, we track causal inclusion relationships and quantify different scenarios. We observe sites including libraries in ad hoc and often transitive ways, which can lead to different versions of the same library being loaded into the same document at the same time. Furthermore, we find that libraries included transitively, or via ad and tracking code, are more likely to be vulnerable. This demonstrates that not only website administrators, but also the dynamic architecture and developers of third-party services are to blame for the Web's poor state of library management. The results of our work underline the need for more thorough approaches to dependency management, code maintenance and third-party code inclusion on the Web.

Sajjad Arshad, Seyed Ali Mirheidari, Tobias Lauinger et al.

Relative Path Overwrite (RPO) is a recent technique to inject style directives into sites even when no style sink or markup injection vulnerability is present. It exploits differences in how browsers and web servers interpret relative paths (i.e., path confusion) to make a HTML page reference itself as a stylesheet; a simple text injection vulnerability along with browsers' leniency in parsing CSS resources results in an attacker's ability to inject style directives that will be interpreted by the browser. Even though style injection may appear less serious a threat than script injection, it has been shown that it enables a range of attacks, including secret exfiltration. In this paper, we present the first large-scale study of the Web to measure the prevalence and significance of style injection using RPO. Our work shows that around 9% of the sites in the Alexa Top 10,000 contain at least one vulnerable page, out of which more than one third can be exploited. We analyze in detail various impediments to successful exploitation, and make recommendations for remediation. In contrast to script injection, relatively simple countermeasures exist to mitigate style injection. However, there appears to be little awareness of this attack vector as evidenced by a range of popular Content Management Systems (CMSes) that we found to be exploitable.

Reza Mirzazade Farkhani, Saman Jafari, Sajjad Arshad et al.

Control flow integrity (CFI) has received significant attention in the community to combat control hijacking attacks in the presence of memory corruption vulnerabilities. The challenges in creating a practical CFI has resulted in the development of a new type of CFI based on runtime type checking (RTC). RTC-based CFI has been implemented in a number of recent practical efforts such as GRSecurity Reuse Attack Protector (RAP) and LLVM-CFI. While there has been a number of previous efforts that studied the strengths and limitations of other types of CFI techniques, little has been done to evaluate the RTC-based CFI. In this work, we study the effectiveness of RTC from the security and practicality aspects. From the security perspective, we observe that type collisions are abundant in sufficiently large code bases but exploiting them to build a functional attack is not straightforward. Then we show how an attacker can successfully bypass RTC techniques using a variant of ROP attacks that respect type checking (called TROP) and also built two proof-of-concept exploits, one against Nginx web server and the other against Exim mail server. We also discuss practical challenges of implementing RTC. Our findings suggest that while RTC is more practical for applying CFI to large code bases, its policy is not strong enough when facing a motivated attacker.