Achieving reproducibility, quantity, and diversity in vulnerability datasets has long been viewed as an inherent three-way trade-off, where improving one dimension often comes at the cost of the others. In practice, reproducibility has been the dimension most often neglected. This has limited what can be automatically extracted from historical bug datasets, and has reduced their utility for downstream security research. In this work, we propose a method to produce a new security dataset which ensures reproducibility for diverse vulnerabilities at scale by identifying the key obstacles to large-scale bug reproduction and addressing them with general solutions. Using this method, we introduce full reproducibility to the largest open source software vulnerability dataset (OSS-Fuzz) and construct the ARVO dataset (an Atlas of Reproducible Vulnerabilities in Open-source software). ARVO is a large-scale dataset consisting of over 6,100 real-world vulnerabilities across 311 projects. Focusing on reproducibility, ARVO differs from existing datasets by providing each vulnerability in a form that can be consistently rebuilt, triggered, and analyzed across versions. Reproducibility also enables automatic identification of the corresponding patch for each vulnerability and supports direct interaction with vulnerabilities after code changes, capabilities that existing large-scale datasets do not provide. In our evaluation, ARVO successfully reproduces 81% of vulnerabilities and achieves 89.4% accuracy on the located patches. We also discuss ARVO's influence on both upstream practices and downstream security research.

Although language model (LM) agents have demonstrated increased performance in multiple domains, including coding and web-browsing, their success in cybersecurity has been limited. We present *EnIGMA*, an LM agent for autonomously solving Capture The Flag (CTF) challenges. We introduce new tools and interfaces to improve the agent's ability to find and exploit security vulnerabilities, focusing on interactive terminal programs. These novel *Interactive Agent Tools* enable LM agents, for the first time, to run interactive utilities, such as a debugger and a server connection tool, which are essential for solving these challenges. Empirical analysis on 390 CTF challenges across four benchmarks demonstrate that these new tools and interfaces substantially improve our agent's performance, achieving state-of-the-art results on NYU CTF, Intercode-CTF, and CyBench. Finally, we analyze data leakage, developing new methods to quantify it and identifying a new phenomenon we term *soliloquizing*, where the model self-generates hallucinated observations without interacting with the environment.

The rapid advancement of large language models (LLMs) has significantly improved code completion tasks, yet the trade-off between accuracy and computational cost remains a critical challenge. While using larger models and incorporating inference-time self-testing algorithms can significantly improve output accuracy, they incur substantial computational expenses at the same time. Furthermore, servers in real-world scenarios usually have a dynamic preference on the cost-accuracy tradeoff, depending on the budget, bandwidth, the concurrent user volume, and users’ sensitivity to wrong answers. In this work, we introduce a novel framework combining model cascading and inference-time self-feedback algorithms to find multiple near-optimal self-testing options on the cost-accuracy tradeoff in LLM-based code generation. Our approach leverages self-generated tests to both enhance accuracy and evaluate model cascading decisions. As a blackbox inference-time method, it requires no access to internal model parameters. We further propose a threshold-based algorithm to determine when to deploy larger models and a heuristic to optimize the number of solutions, test cases, and test lines generated per model, based on budget constraints. Experimental results show that our cascading approach reduces costs by an average of 26%, and up to 70% in the best case, across various model families and datasets, while maintaining or improving accuracy in natural language generation tasks compared to both random and optimal single-model self-testing schemes. To our knowledge, this is the first work to provide a series of choices for optimizing the cost-accuracy trade-off in LLM code generation with self-testing.

The security landscape of software systems has witnessed considerable advancements through dynamic testing methodologies, especially fuzzing. Traditionally, fuzzing involves a sequential, cyclic process where software is tested to identify crashes. These crashes are then triaged and patched, leading to subsequent cycles that uncover further vulnerabilities. While effective, this method is not efficient as each cycle potentially reveals new issues previously obscured by earlier crashes, thus resulting in vulnerabilities being discovered sequentially.

Large Language Models (LLMs) are being deployed across various domains today. However, their capacity to solve Capture the Flag (CTF) challenges in cybersecurity has not been thoroughly evaluated. To address this, we develop a novel method to assess LLMs in solving CTF challenges by creating a scalable, open-source benchmark database specifically designed for these applications. This database includes metadata for LLM testing and adaptive learning, compiling a diverse range of CTF challenges from popular competitions. Utilizing the advanced function calling capabilities of LLMs, we build a fully automated system with an enhanced workflow and support for external tool calls. Our benchmark dataset and automated framework allow us to evaluate the performance of five LLMs, encompassing both black-box and open-source models. This work lays the foundation for future research into improving the efficiency of LLMs in interactive cybersecurity tasks and automated task planning. By providing a specialized benchmark, our project offers an ideal platform for developing, testing, and refining LLM-based approaches to vulnerability detection and resolution. Evaluating LLMs on these challenges and comparing with human performance yields insights into their potential for AI-driven cybersecurity solutions to perform real-world threat management. We make our benchmark dataset open source to public https://github.com/NYU-LLM-CTF/NYUCTFBench along with our playground automated framework https://github.com/NYU-LLM-CTF/llmctfautomation.

The security of computer systems typically relies on a hardware root of trust. As vulnerabilities in hardware can have severe implications on a system, there is a need for techniques to support security verification activities. Assertion-based verification is a popular verification technique that involves capturing design intent in a set of assertions that can be used in formal verification or testing-based checking. However, writing security-centric assertions is a challenging task. In this work, we investigate the use of emerging large language models (LLMs) for code generation in hardware assertion generation for security, where primarily natural language prompts, such as those one would see as code comments in assertion files, are used to produce SystemVerilog assertions. We focus our attention on a popular LLM and characterize its ability to write assertions out of the box, given varying levels of detail in the prompt. We design an evaluation framework that generates a variety of prompts, and we create a benchmark suite comprising real-world hardware designs and corresponding golden reference assertions that we want to generate with the LLM.

In this study, we explore the capability of Large Language Models (LLMs) to automate hardware design by automatically completing partial Verilog code, a common language for designing and modeling digital systems. We fine-tune pre-existing LLMs on Verilog datasets compiled from GitHub and Verilog textbooks. We evaluate the functional correctness of the generated Verilog code using a specially designed test suite, featuring a custom problem set and testing benches. Here, our fine-tuned open-source CodeGen-16B model outperforms the commercial state-of-the-art GPT-3.5-turbo model with a 1.1% overall increase. Upon testing with a more diverse and complex problem set, we find that the fine-tuned model shows competitive performance against state-of-the-art gpt-3.5-turbo, excelling in certain scenarios. Notably, it demonstrates a 41% improvement in generating syntactically correct Verilog code across various problem categories compared to its pre-trained counterpart, highlighting the potential of smaller, in-house LLMs in hardware design automation. We release our training/evaluation scripts and LLM checkpoints as open-source contributions.

Automating hardware design could obviate a signif-icant amount of human error from the engineering process and lead to fewer errors. Verilog is a popular hardware description language to model and design digital systems, thus generating Verilog code is a critical first step. Emerging large language models (LLMs) are able to write high-quality code in other programming languages. In this paper, we characterize the ability of LLMs to generate useful Verilog. For this, we fine-tune pre-trained LLMs on Verilog datasets collected from GitHub and Verilog textbooks. We construct an evaluation framework comprising test-benches for functional analysis and a flow to test the syntax of Verilog code generated in response to problems of varying difficulty. Our findings show that across our problem scenarios, the fine-tuning results in LLMs more capable of producing syntactically correct code (25.9% overall). Further, when analyzing functional correctness, a fine-tuned open-source CodeGen LLM can outperform the state-of-the-art commercial Codex LLM (6.5% overall). We release our training/evaluation scripts and LLM checkpoints as open source contributions.

Fuzz testing is often automated, but also frequently augmented by experts who insert themselves into the workflow in a greedy search for bugs. In this paper, we propose Homo in Machina, or HM-fuzzing, in which analyses guide the manual efforts, maximizing benefit. As one example of this paradigm, we introduce compartment analysis. Compartment analysis uses a whole-program dominator analysis to estimate the utility of reaching new code, and combines this with a dynamic analysis indicating drastically under-covered edges guarding that code. This results in a prioritized list of compartments, i.e., large, uncovered parts of the pro-gram semantically partitioned and largely unreachable given the current corpus of inputs under consideration. A human can use this categorization and ranking of compartments directly to focus manual effort, finding or fashioning inputs that make the compartments available for future fuzzing. We evaluate the effect of compartment analysis on seven projects within the OSS-Fuzz corpus where we see coverage improvements over AFL++ as high as 94%, with a median of 13%. We further observe that the determination of compartments is highly stable and thus can be done early in a fuzzing campaign, maximizing the potential for impact.

Human developers can produce code with cybersecurity bugs. Can emerging ‘smart’ code completion tools help repair those bugs? In this work, we examine the use of large language models (LLMs) for code (such as OpenAI’s Codex and AI21’s Jurassic J-1) for zero-shot vulnerability repair. We investigate challenges in the design of prompts that coax LLMs into generating repaired versions of insecure code. This is difficult due to the numerous ways to phrase key information— both semantically and syntactically—with natural languages. We perform a large scale study of five commercially available, black-box, "off-the-shelf" LLMs, as well as an open-source model and our own locally-trained model, on a mix of synthetic, hand-crafted, and real-world security bug scenarios. Our experiments demonstrate that while the approach has promise (the LLMs could collectively repair 100% of our synthetically generated and hand-crafted scenarios), a qualitative evaluation of the model’s performance over a corpus of historical real-world examples highlights challenges in generating functionally correct code.

Human developers can produce code with cybersecurity bugs. Can emerging ‘smart’ code completion tools help repair those bugs? In this work, we examine the use of large language models (LLMs) for code (such as OpenAI’s Codex and AI21’s Jurassic J-1) for zero-shot vulnerability repair. We investigate challenges in the design of prompts that coax LLMs into generating repaired versions of insecure code. This is difficult due to the numerous ways to phrase key information— both semantically and syntactically—with natural languages. We perform a large scale study of five commercially available, black-box, "off-the-shelf" LLMs, as well as an open-source model and our own locally-trained model, on a mix of synthetic, hand-crafted, and real-world security bug scenarios. Our experiments demonstrate that while the approach has promise (the LLMs could collectively repair 100% of our synthetically generated and hand-crafted scenarios), a qualitative evaluation of the model’s performance over a corpus of historical real-world examples highlights challenges in generating functionally correct code.

Sophisticated attackers find bugs in software, evaluate their exploitability, and then create and launch exploits for bugs found to be exploitable. Most efforts to secure software attempt either to eliminate bugs or to add mitigations that make exploitation more difficult. In this paper, we propose a new defensive technique called chaff bugs, which instead targets the bug discovery and exploit creation stages of this process. Rather than eliminating bugs, we instead add large numbers of bugs that are non-exploitable. Attackers who attempt to find and exploit bugs in software will, with high probability, find an intentionally placed non-exploitable bug and waste precious resources in trying to build a working exploit. In a prototype, we demonstrate two strategies for ensuring non-exploitability for memory safety bugs in C/C++ programs and use them to automatically add thousands of non-exploitable bugs to real-world software such as nginx and libFLAC; we show that the functionality of the software is not impaired and demonstrate that our bugs look exploitable to current triage tools. We believe that chaff bugs can serve as an effective deterrent against both human attackers and automated bug-finding tools.

Embedded and IoT devices often come with a wide range of hardware functionality, but any particular end user may only use some small subset of these features. However, even unused hardware features are accompanied by potentially buggy driver code, which increases the attack surface of the device. In this paper, we introduce IRQDebloat, a system for disabling unwanted hardware features through automated firmware rewriting. Building on the insight that external inputs to the system are typically delivered through interrupt requests (IRQs), IRQDebloat systematically explores the interrupt handling code in the target firmware, identifies the handler function for each peripheral, and finally rewrites target firmware to disable the handlers that correspond to undesired hardware features. In our experiments we demonstrate IRQDebloat’s effectiveness and generality by identifying IRQ handlers across four different operating systems (Linux, FreeBSD, VxWorks, and RiscOS) and seven different embedded platforms, and disabling selected peripherals on real-world hardware (a Raspberry Pi and a Valve Steam Link). On the Steam Link, we survey the attack surface and find that disabling selected peripherals could block up to 44 CVEs found in the Linux kernel over the past five years.

There is burgeoning interest in designing AI-based systems to assist humans in designing computing systems, including tools that automatically generate computer code. The most notable of these comes in the form of the first self-described ‘AI pair programmer’, GitHub Copilot, which is a language model trained over open-source GitHub code. However, code often contains bugs—and so, given the vast quantity of unvetted code that Copilot has processed, it is certain that the language model will have learned from exploitable, buggy code. This raises concerns on the security of Copilot’s code contributions. In this work, we systematically investigate the prevalence and conditions that can cause GitHub Copilot to recommend insecure code. To perform this analysis we prompt Copilot to generate code in scenarios relevant to high-risk cybersecurity weaknesses, e.g. those from MITRE’s “Top 25” Common Weakness Enumeration (CWE) list. We explore Copilot’s performance on three distinct code generation axes—examining how it performs given diversity of weaknesses, diversity of prompts, and diversity of domains. In total, we produce 89 different scenarios for Copilot to complete, producing 1,689 programs. Of these, we found approximately 40% to be vulnerable.

Automated bug-finding tools such as KLEE have achieved mainstream success over the last decade, and have proved capable of finding deep bugs even in programs that have received significant manual testing. Some recent works have demonstrated techniques for finding bugs in these bug-finding tools themselves; however, it remains unclear whether these correctness issues have any practical impact on their ability to uncover serious bugs. In this paper, we study this issue by conducting experiments with KLEE 1.4 and 2.2 on several corpora of memory safety bugs. Using automated bug injection, we can automatically find false negatives (i.e., bugs missed by KLEE); moreover, because the bugs we inject come with triggering inputs, we can then use concolic execution to tell which bugs were missed due path explosion and which are caused by soundness issues in KLEE. Our evaluation uncovers several sources of unsoundness, including a limitation in how KLEE detects memory errors, mismatches in the modeling of the C standard library, lack of support for floating point and C++, and issues with calls to external functions. Our results suggest that bug injection and other synthetic corpora can help highlight implementation issues in current tools and illuminate directions for future research in automated software engineering.

Fuzz testing has been used to find bugs in programs since the 1990s, but despite decades of dedicated research, there is still no consensus on which fuzzing techniques work best. One reason for this is the paucity of ground truth: bugs in real programs with known root causes and triggering inputs are difficult to collect at a meaningful scale. Bug injection technologies that add synthetic bugs into real programs seem to offer a solution, but the differences in finding these synthetic bugs versus organic bugs have not previously been explored at a large scale. Using over 80 years of CPU time, we ran eight fuzzers across 20 targets from the Rode0day bug-finding competition and the LAVA-M corpus. Experiments were standardized with respect to compute resources and metrics gathered. These experiments show differences in fuzzer performance as well as the impact of various configuration options. For instance, it is clear that integrating symbolic execution with mutational fuzzing is very effective and that using dictionaries improves performance. Other conclusions are less clear-cut; for example, no one fuzzer beat all others on all tests. It is noteworthy that no fuzzer found any organic bugs (i.e., one reported in a CVE), despite 50 such bugs being available for discovery in the fuzzing corpus. A close analysis of results revealed a possible explanation: a dramatic difference between where synthetic and organic bugs live with respect to the "main path" discovered by fuzzers. We find that recent updates to bug injection systems have made synthetic bugs more difficult to discover, but they are still significantly easier to find than organic bugs in our target programs. Finally, this study identifies flaws in bug injection techniques and suggests a number of axes along which synthetic bugs should be improved.

Closely monitoring the behavior of a software system during its execution enables developers and analysts to observe, and ultimately understand, how it works. This kind of dynamic analysis can be instrumental to reverse engineering, vulnerability discovery, exploit development, and debugging. While these analyses are typically well-supported for homogeneous desktop platforms (e.g., x86 desktop PCs), they can rarely be applied in the heterogeneous world of embedded systems. One approach to enable dynamic analyses of embedded systems is to move software stacks from physical systems into virtual environments that sufficiently model hardware behavior. This process which we call "rehosting" poses a significant research challenge with major implications for security analyses. Although rehosting has traditionally been an unscientific and ad-hoc endeavor undertaken by domain experts with varying time and resources at their disposal, researchers are beginning to address rehosting challenges systematically and in earnest. In this paper, we establish that emulation is insufficient to conduct large-scale dynamic analysis of real-world hardware systems and present rehosting as a firmware-centric alternative. Furthermore, we taxonomize preliminary rehosting efforts, identify the fundamental components of the rehosting process, and propose directions for future research.

Use-after-free (UAF) vulnerabilities, in which dangling pointers remain after memory is released, remain a persistent problem for applications written in C and C++. In order to protect legacy code, prior work has attempted to track pointer propagation and invalidate dangling pointers at deallocation time, but this work has gaps in coverage, as it lacks support for tracking program variables promoted to CPU registers. Moreover, we find that these gaps can significantly hamper detection of UAF bugs: in a preliminary study with OSS-Fuzz, we found that more than half of the UAFs in real-world programs we examined (10/19) could not be detected by prior systems due to register promotion. In this paper, we introduce HeapExpo, a new system that fills this gap in coverage by parsimoniously identifying potential dangling pointer variables that may be lifted into registers by the compiler and marking them as volatile. In our experiments, we find that HeapExpo effectively detects UAFs missed by other systems with an overhead of 35% on the majority of SPEC CPU2006 and 66% when including two benchmarks that have high amounts of pointer propagation.

Mobile applications (apps) have exploded in popularity, with billions of smartphone users using millions of apps available through markets such as the Google Play Store or the Apple App Store. While these apps have rich and useful functionality that is publicly exposed to end users, they also contain hidden behaviors that are not disclosed, such as backdoors and blacklists designed to block unwanted content. In this paper, we show that the input validation behavior-the way the mobile apps process and respond to data entered by users-can serve as a powerful tool for uncovering such hidden functionality. We therefore have developed a tool, InputScope, that automatically detects both the execution context of user input validation and also the content involved in the validation, to automatically expose the secrets of interest. We have tested InputScope with over 150,000 mobile apps, including popular apps from major app stores and preinstalled apps shipped with the phone, and found 12,706 mobile apps with backdoor secrets and 4,028 mobile apps containing blacklist secrets.

The strength of an anonymity system depends on the number of users. Therefore, User eXperience (UX) and usability of these systems is of critical importance for boosting adoption and use. To this end, we carried out a study with 19 non-expert participants to investigate how users experience routine Web browsing via the Tor Browser, focusing particularly on encountered problems and frustrations. Using a mixed-methods quantitative and qualitative approach to study one week of naturalistic use of the Tor Browser, we uncovered a variety of UX issues, such as broken Web sites, latency, lack of common browsing conveniences, differential treatment of Tor traffic, incorrect geolocation, operational opacity, etc. We applied this insight to suggest a number of UX improvements that could mitigate the issues and reduce user frustration when using the Tor Browser.

In spite of decades of research in bug detection tools, there is a surprising dearth of ground-truth corpora that can be used to evaluate the efficacy of such tools. Recently, systems such as LAVA and EvilCoder have been proposed to automatically inject bugs into software to quickly generate large bug corpora, but the bugs created so far differ from naturally occurring bugs in a number of ways. In this work, we propose a new automated bug injection system, Apocalypse, that uses formal techniques—symbolic execution, constraint-based program synthesis and model counting—to automatically inject fair (can potentially be discovered by current bug-detection tools), deep (requiring a long sequence of dependencies to be satisfied to fire), uncorrelated (each bug behaving independent of others), reproducible (a trigger input being available) and rare (can be triggered by only a few program inputs) bugs in large software code bases. In our evaluation, we inject bugs into thirty Coreutils programs as well as the TCAS test suite. We find that bugs synthesized by Apocalypse are highly realistic under a variety of metrics, that they do not favor a particular bug-finding strategy (unlike bugs produced by LAVA), and that they are more difficult to find than manually injected bugs, requiring up around 240× more tests to discover with a state-of-the-art symbolic execution tool.

Work on automating vulnerability discovery has long been hampered by a shortage of ground-truth corpora with which to evaluate tools and techniques. This lack of ground truth prevents authors and users of tools alike from being able to measure such fundamental quantities as miss and false alarm rates. In this paper, we present LAVA, a novel dynamic taint analysis-based technique for producing ground-truth corpora by quickly and automatically injecting large numbers of realistic bugs into program source code. Every LAVA bug is accompanied by an input that triggers it whereas normal inputs are extremely unlikely to do so. These vulnerabilities are synthetic but, we argue, still realistic, in the sense that they are embedded deep within programs and are triggered by real inputs. Using LAVA, we have injected thousands of bugs into eight real-world programs, including bash, tshark, and the GNU coreutils. In a preliminary evaluation, we found that a prominent fuzzer and a symbolic execution-based bug finder were able to locate some but not all LAVA-injected bugs, and that interesting patterns and pathologies were already apparent in their performance. Our work forms the basis of an approach for generating large ground-truth vulnerability corpora on demand, enabling rigorous tool evaluation and providing a high-quality target for tool developers.

The ability to introspect into the behavior of software at runtime is crucial for many security-related tasks, such as virtual machine-based intrusion detection and low-artifact malware analysis. Although some progress has been made in this task by automatically creating programs that can passively retrieve kernel-level information, two key challenges remain. First, it is currently difficult to extract useful information from user-level applications, such as web browsers. Second, discovering points within the OS and applications to hook for active monitoring is still an entirely manual process. In this paper we propose a set of techniques to mine the memory accesses made by an operating system and its applications to locate useful places to deploy active monitoring, which we call tap points. We demonstrate the efficacy of our techniques by finding tap points for useful introspection tasks such as finding SSL keys and monitoring web browser activity on five different operating systems (Windows 7, Linux, FreeBSD, Minix and Haiku) and two processor architectures (ARM and x86).

Introspection has featured prominently in many recent security solutions, such as virtual machine-based intrusion detection, forensic memory analysis, and low-artifact malware analysis. Widespread adoption of these approaches, however, has been hampered by the semantic gap: in order to extract meaningful information about the current state of a virtual machine, detailed knowledge of the guest operating system's inner workings is required. In this paper, we present a novel approach for automatically creating introspection tools for security applications with minimal human effort. By analyzing dynamic traces of small, in-guest programs that compute the desired introspection information, we can produce new programs that retrieve the same information from outside the guest virtual machine. We demonstrate the efficacy of our techniques by automatically generating 17 programs that retrieve security information across 3 different operating systems, and show that their functionality is unaffected by the compromise of the guest system. Our technique allows introspection tools to be effortlessly generated for multiple platforms, and enables the development of rich introspection-based security applications.

Kernel-mode rootkits hide objects such as processes and threads using a technique known as Direct Kernel Object Manipulation (DKOM). Many forensic analysis tools attempt to detect these hidden objects by scanning kernel memory with handmade signatures; however, such signatures are brittle and rely on non-essential features of these data structures, making them easy to evade. In this paper, we present an automated mechanism for generating signatures for kernel data structures and show that these signatures are robust: attempts to evade the signature by modifying the structure contents will cause the OS to consider the object invalid. Using dynamic analysis, we profile the target data structure to determine commonly used fields, and we then fuzz those fields to determine which are essential to the correct operation of the OS. These fields form the basis of a signature for the data structure. In our experiments, our new signature matched the accuracy of existing scanners for traditional malware and found processes hidden with our prototype rootkit that all current signatures missed. Our techniques significantly increase the difficulty of hiding objects from signature scanning.