Antonio Bianchi

As users increasingly turn to large language model (LLM) based web agents to automate online tasks, agents may encounter dark patterns: deceptive user interface designs that manipulate users into making unintended decisions. Although dark patterns primarily target human users, their potentially harmful impacts on LLM-based generalist web agents remain unexplored. In this paper, we present the first study that investigates the impact of dark patterns on the decision-making process of LLM-based generalist web agents. To achieve this, we introduce LiteAgent, a lightweight framework that automatically prompts agents to execute tasks while capturing comprehensive logs and screen-recordings of their interactions. We also present TrickyArena, a controlled environment comprising web applications from domains such as e-commerce, streaming services, and news platforms, each containing diverse and realistic dark patterns that can be selectively enabled or disabled. Using LiteAgent and TrickyArena, we conduct multiple experiments to assess the impact of both individual and combined dark patterns on web agent behavior. We evaluate six popular LLM-based generalist web agents across three LLMs and discover that when there is a single dark pattern present, agents are susceptible to it an average of 41% of the time. We also find that modifying dark pattern UI attributes through visual design changes or HTML code adjustments and introducing multiple dark patterns simultaneously can influence agent susceptibility. This study emphasizes the need for holistic defense mechanisms in web agents, encompassing both agent-specific protections and broader web safety measures.

Modern computing systems remain rife with software vulnerabilities. Engineers apply many means to detect them, of which dynamic testing is one of the most common and effective. However, most dynamic testing techniques follow a top-down paradigm, and struggle to reach and exercise functions deep within the call graph. While recent works have proposed Bottom-Up approaches to address these limitations, they face challenges with false positives and generating valid inputs that adhere to the context of the entire program. In this work, we introduce a new paradigm that we call Reactive Bottom-Up Testing. Our insight is that function-level testing is necessary but not sufficient for the validation of vulnerabilities in functions. What we need is a systematic approach that not only tests functions in isolation but also validates their behavior within the broader program context, ensuring that detected vulnerabilities are both reachable and triggerable. We develop a three-stage bottom-up testing scheme: (1) identify likely-vulnerable functions and generate type- and context-aware harnesses; (2) fuzz to find crashes and extract input constraints via symbolic execution; (3) verify crashes by combining constraints to remove false positives. We implemented an automated prototype, which we call Griller. We evaluated Griller in a controlled setting using a benchmark of 48 known vulnerabilities across 5 open-source projects, where we successfully detected 28 known vulnerabilities. Additionally, we evaluated Griller on several real-world applications such as Pacman, and it discovered 6 previously unknown vulnerabilities. Our findings suggest that Reactive Bottom-Up Testing can significantly enhance the detection of vulnerabilities in complex systems, paving the way for more robust security practices.

Developers are increasingly integrating Language Models (LMs) into their mobile apps to provide features such as chat-based assistants. To prevent LM misuse, they impose various restrictions, including limits on the number of queries, input length, and allowed topics. However, if the LM integration is insecure, attackers can bypass these restrictions and gain unrestricted access to the LM, potentially harming developers' reputations and leading to significant financial losses. This paper presents the first systematic study of insecure usage of LMs by Android apps. We first manually analyze a preliminary dataset of apps to investigate LM integration methods, construct a taxonomy that categorizes the LM usage restrictions implemented by the apps, and determine how to bypass them. Alarmingly, we can bypass restrictions in 127 out of 181 apps. Then, we develop LM-Scout, a fully automated tool to detect on a large-scale vulnerable usage of LMs in 2,950 mobile apps. LM-Scout shows that, in many cases (i.e., 120 apps), it is possible to find and exploit such security issues automatically. Finally, we identify the root causes for the identified issues and offer recommendations for secure LM integration.

In Voice Assistant (VA) platforms, when users add devices to their accounts and give voice commands, complex interactions occur between the devices, skills, VA clouds, and vendor clouds. These interactions are governed by the device management capabilities (DMC) of VA platforms, which rely on device names, types, and associated skills in the user account. Prior work studied vulnerabilities in specific VA components, such as hidden voice commands and bypassing skill vetting. However, the security and privacy implications of device management flaws have largely been unexplored. In this paper, we introduce DMC-Xplorer, a testing framework for the automated discovery of VA device management flaws. We first introduce VA description language (VDL), a new domain-specific language to create VA environments for testing, using VA and skill developer APIs. DMC-Xplorer then selects VA parameters (device names, types, vendors, actions, and skills) in a combinatorial approach and creates VA environments with VDL. It issues real voice commands to the environment via developer APIs and logs event traces. It validates the traces against three formal security properties that define the secure operation of VA platforms. Lastly, DMC-Xplorer identifies the root cause of property violations through intervention analysis to identify VA device management flaws. We exercised DMC-Xplorer on Amazon Alexa and Google Home and discovered two design flaws that can be exploited to launch four attacks. We show that malicious skills with default permissions can eavesdrop on privacy-sensitive device states, prevent users from controlling their devices, and disrupt the services on the VA cloud.

There is a growing interest in integrating Large Language Models (LLMs) with autonomous driving (AD) systems. However, AD systems are vulnerable to attacks against their object detection and tracking (ODT) functions. Unfortunately, our evaluation of four recent LLM agents against ODT attacks shows that the attacks are 63.26% successful in causing them to crash or violate traffic rules due to (1) misleading memory modules that provide past experiences for decision making, (2) limitations of prompts in identifying inconsistencies, and (3) reliance on ground truth perception data. In this paper, we introduce Hudson, a driving reasoning agent that extends prior LLM-based driving systems to enable safer decision making during perception attacks while maintaining effectiveness under benign conditions. Hudson achieves this by first instrumenting the AD software to collect real-time perception results and contextual information from the driving scene. This data is then formalized into a domain-specific language (DSL). To guide the LLM in detecting and making safe control decisions during ODT attacks, Hudson translates the DSL into natural language, along with a list of custom attack detection instructions. Following query execution, Hudson analyzes the LLM's control decision to understand its causal reasoning process. We evaluate the effectiveness of Hudson using a proprietary LLM (GPT-4) and two open-source LLMs (Llama and Gemma) in various adversarial driving scenarios. GPT-4, Llama, and Gemma achieve, on average, an attack detection accuracy of 83. 3%, 63. 6%, and 73. 6%. Consequently, they make safe control decisions in 86.4%, 73.9%, and 80% of the attacks. Our results, following the growing interest in integrating LLMs into AD systems, highlight the strengths of LLMs and their potential to detect and mitigate ODT attacks.

With the continuous rise in the popularity of Android mobile devices, automated testing of apps has become more important than ever. Android apps are event-driven programs. Unfortunately, generating all possible types of events by interacting with the app's interface is challenging for an automated testing approach. Callback-driven testing eliminates the need for event generation by directly invoking app callbacks. However, existing callback-driven testing techniques assume prior knowledge of Android callbacks, and they rely on a human expert, who is familiar with the Android API, to write stub code that prepares callback arguments before invocation. Since the Android API is huge and keeps evolving, prior techniques could only support a small fraction of callbacks present in the Android framework. In this work, we introduce Columbus, a callback-driven testing technique that employs two strategies to eliminate the need for human involvement: (i) it automatically identifies callbacks by simultaneously analyzing both the Android framework and the app under test, and (ii) it uses a combination of under-constrained symbolic execution (primitive arguments), and type-guided dynamic heap introspection (object arguments) to generate valid and effective inputs. Lastly, Columbus integrates two novel feedback mechanisms -- data dependency and crash-guidance, during testing to increase the likelihood of triggering crashes, and maximizing coverage. In our evaluation, Columbus outperforms state-of-the-art model-driven, checkpoint-based, and callback-driven testing tools both in terms of crashes and coverage.