In Google’s FuzzBench platform, we find that the outcome of coverage-based evaluation more strongly agrees with the outcome of a bug-based evaluation than an independent bug-based evaluation itself. Recently, B'ohme et al. found that despite a very strong correlation between coverage achieved and bugs found, there is no strong agreement between the outcome of a coverage- and a bug-based evaluation: The fuzzer best at achieving coverage may be the worst at finding bugs. However, in trying to explain this moderate agreement, we wondered whether the outcome of bug-based benchmarking itself is perhaps much more “noisy” and turned to applied statistics to develop the tools necessary to investigate our hypothesis. In this paper, we call this degree of “noisiness” of a benchmarking outcome the concordance of the benchmarking procedure and quantify it using a measure of statistical reliability widely used in psychology, called mean split-half reliability , i.e., the expected agreement on the benchmark outcome between two random halves of the benchmarking suite. In our experiments with FuzzBench and Magma, we find that the concordance of coverage-based benchmarking is consistently strong while that of bug-based benchmarking is weak on FuzzBench and moderate on Magma. In contrast to FuzzBench, for the Magma benchmark suite (which was designed for bug-based evaluation) a coverage-based evaluation does not predict the outcome of a bug-based evaluation better than an independent bug-based evaluation. Moreover, to demonstrate the utility of concordance also for developers of benchmarking suites, we investigate concordance as a measure of benchmarking efficiency, as in green fuzzer benchmarking. We empirically confirm that the resources of a procedure with higher concordance can be reduced more substantially (in terms of campaign length or benchmark sampling size) while maintaining a similar benchmark outcome as a procedure with lower concordance. We report the corresponding savings in terms of carbon emissions.

Large language models (LLMs) introduce new security risks, but there are few comprehensive evaluation suites to measure and reduce these risks. We present BenchmarkName, a novel benchmark to quantify LLM security risks and capabilities. We introduce two new areas for testing: prompt injection and code interpreter abuse. We evaluated multiple state-of-the-art (SOTA) LLMs, including GPT-4, Mistral, Meta Llama 3 70B-Instruct, and Code Llama. Our results show that conditioning away risk of attack remains an unsolved problem; for example, all tested models showed between 26% and 41% successful prompt injection tests. We further introduce the safety-utility tradeoff: conditioning an LLM to reject unsafe prompts can cause the LLM to falsely reject answering benign prompts, which lowers utility. We propose quantifying this tradeoff using False Refusal Rate (FRR). As an illustration, we introduce a novel test set to quantify FRR for cyberattack helpfulness risk. We find many LLMs able to successfully comply with "borderline" benign requests while still rejecting most unsafe requests. Finally, we quantify the utility of LLMs for automating a core cybersecurity task, that of exploiting software vulnerabilities. This is important because the offensive capabilities of LLMs are of intense interest; we quantify this by creating novel test sets for four representative problems. We find that models with coding capabilities perform better than those without, but that further work is needed for LLMs to become proficient at exploit generation. Our code is open source and can be used to evaluate other LLMs.

This paper presents CyberSecEval, a comprehensive benchmark developed to help bolster the cybersecurity of Large Language Models (LLMs) employed as coding assistants. As what we believe to be the most extensive unified cybersecurity safety benchmark to date, CyberSecEval provides a thorough evaluation of LLMs in two crucial security domains: their propensity to generate insecure code and their level of compliance when asked to assist in cyberattacks. Through a case study involving seven models from the Llama 2, Code Llama, and OpenAI GPT large language model families, CyberSecEval effectively pinpointed key cybersecurity risks. More importantly, it offered practical insights for refining these models. A significant observation from the study was the tendency of more advanced models to suggest insecure code, highlighting the critical need for integrating security considerations in the development of sophisticated LLMs. CyberSecEval, with its automated test case generation and evaluation pipeline covers a broad scope and equips LLM designers and researchers with a tool to broadly measure and enhance the cybersecurity safety properties of LLMs, contributing to the development of more secure AI systems.

Coverage-guided fuzz testing ("fuzzing") has become mainstream and we have observed lots of progress in this research area recently. However, it is still challenging to efficiently test network services with existing coverage-guided fuzzing methods. In this paper, we introduce the design and implementation of Nyx-Net, a novel snapshot-based fuzzing approach that can successfully fuzz a wide range of targets spanning servers, clients, games, and even Firefox's Inter-Process Communication (IPC) interface. Compared to state-of-the-art methods, Nyx-Net improves test throughput by up to 300x and coverage found by up to 70%. Additionally, Nyx-Net is able to find crashes in two of ProFuzzBench's targets that no other fuzzer found previously. When using Nyx-Net to play the game Super Mario, Nyx-Net shows speedups of 10--30x compared to existing work. Moreover, Nyx-Net is able to find previously unknown bugs in servers such as Lighttpd, clients such as MySQL client, and even Firefox's IPC mechanism---demonstrating the strength and versatility of the proposed approach. Lastly, our prototype implementation was awarded a $20.000 bug bounty for enabling fuzzing on previously unfuzzable code in Firefox and solving a long-standing problem at Mozilla.

Coverage-guided fuzz testing (fuzzing) has become mainstream and we have observed lots of progress in this research area recently. However, it is still challenging to efficiently test network services with existing coverage-guided fuzzing methods. In this paper, we introduce the design and implementation of Nyx-Net, a novel snapshot-based fuzzing approach that can successfully fuzz a wide range of targets spanning servers, clients, games, and even Firefox's Inter-Process Communication (IPC) interface. Compared to state-of-the-art methods, Nyx-Net improves test throughput by up to 300x and coverage found by up to 70%. Additionally, Nyx-Net is able to find crashes in two of ProFuzzBench's targets that no other fuzzer found previously. When using Nyx-Net to play the game Super Mario, Nyx-Net shows speedups of 10-30x compared to existing work. Under some circumstances, Nyx-Net is even able play faster than light: solving the level takes less wall-clock time than playing the level perfectly even once. Nyx-Net is able to find previously unknown bugs in servers such as Lighttpd, clients such as MySQL client, and even Firefox's IPC mechanism - demonstrating the strength and versatility of the proposed approach. Lastly, our prototype implementation was awarded a $20.000 bug bounty for enabling fuzzing on previously unfuzzable code in Firefox and solving a long-standing problem at Mozilla.

A hypervisor (also know as virtual machine monitor, VMM) enforces the security boundaries between different virtual machines (VMs) running on the same physical machine. A malicious user who is able to run her own kernel on a cloud VM can interact with a large variety of attack surfaces. Exploiting a software fault in any of these surfaces leads to full access to all other VMs that are co-located on the same host. Hence, the efficient detection of hypervisor vulnerabilities is crucial for the security of the modern cloud infrastructure. Recent work showed that blind fuzzing is the most efficient approach to identify security issues in hypervisors, mainly due to an outstandingly high test throughput. In this paper we present the design and implementation of NYX, a highly optimized, coverage-guided hypervisor fuzzer. We show how a fast snapshot restoration mechanism that allows us to reload the system under test thousands of times per second is key to performance. Furthermore, we introduce a novel mutation engine based on custom bytecode programs, encoded as directed acyclic graphs (DAG), and affine types, that enables the required flexibility to express complex interactions. Our evaluation shows that, while NYX has a lower throughput than the state-of-the-art hypervisor fuzzer, it performs competitively on simple targets: NYX typically requires only a few minutes longer to achieve the same test coverage. On complex devices, however, our approach is able to significantly outperform existing works. Moreover, we are able to uncover substantially more bugs: in total, we uncovered 44 new bugs with 22 CVEs requested. Our results demonstrate that coverage guidance is highly valuable, even if a blind fuzzer can be significantly faster.

Software obfuscation is a crucial technology to protect intellectual property and manage digital rights within our society. Despite its huge practical importance, both commercial and academic state-of-the-art obfuscation methods are vulnerable to a plethora of automated deobfuscation attacks, such as symbolic execution, taint analysis, or program synthesis. While several enhanced obfuscation techniques were recently proposed to thwart taint analysis or symbolic execution, they either impose a prohibitive runtime overhead or can be removed in an automated way (e.g., via compiler optimizations). In general, these techniques suffer from focusing on a single attack vector, allowing an attacker to switch to other, more effective techniques, such as program synthesis. In this work, we present Loki, an approach for software obfuscation that is resilient against all known automated deobfuscation attacks. To this end, we use and efficiently combine multiple techniques, including a generic approach to synthesize formally verified expressions of arbitrary complexity. Contrary to state-of-the-art approaches that rely on a few hardcoded generation rules, our expressions are more diverse and harder to pattern match against. Even the most recent state-of-the-art research on Mixed-Boolean Arithmetic (MBA) deobfuscation fails to simplify them. Moreover, Loki protects against previously unaccounted attack vectors such as program synthesis, for which it reduces the success rate to merely 19 that our design incurs significantly less overhead while providing a much stronger protection level compared to existing works.

Virtual machine monitors (VMMs, also called hypervisors) represent a very critical part of a modern software stack: compromising them could allow an attacker to take full control of the whole cloud infrastructure of any cloud provider.Hence their security is critical for many applications, especially in the context of Infrastructure-as-a-Service.In this paper, we present the design and implementation of HYPER-CUBE, a novel fuzzer that aims explicitly at testing hypervisors in an efficient, effective, and precise way.Our approach is based on a custom operating system that implements a custom bytecode interpreter.This high-throughput design for long-running, interactive targets allows us to fuzz a large number of both open source and proprietary hypervisors.In contrast to one-dimensional fuzzers such as AFL, HYPER-CUBE can interact with any number of interfaces in any order.Our evaluation results show that we can find more bugs (over 2×) and coverage (as much as 2×) than state-of-the-art hypervisor fuzzers.In most cases, we were even able to do so using multiple orders of magnitude less time than comparable fuzzers.HYPER-CUBE was also able to rediscover a set of well-known hypervisor vulnerabilities, such as VENOM, in less than five minutes.In total, we found 54 novel bugs, and so far obtained 43 CVEs.Our evaluation results demonstrate that nextgeneration coverage-guided fuzzers should incorporate a higherthroughput design for long-running targets such as hypervisors.

Given the huge success of automated software testing techniques, a large amount of crashes is found in practice. Identifying the root cause of a crash is a time-intensive endeavor, causing a disproportion between finding a crash and fixing the underlying software fault. To address this problem, various approaches have been proposed that rely on techniques such as reverse execution and backward taint analysis. Still, these techniques are either limited to certain fault types or provide an analyst with assembly instructions, but no context information or explanation of the underlying fault. In this paper, we propose an automated analysis approach that does not only identify the root cause of a given crashing input for a binary executable, but also provides the analyst with context information on the erroneous behavior that characterizes crashing inputs. Starting with a single crashing input, we generate a diverse set of similar inputs that either also crash the program or induce benign behavior. We then trace the program's states while executing each found input and generate predicates, i. e., simple Boolean expressions that capture behavioral differences between crashing and non-crashing inputs. A statistical analysis of all predicates allows us to identify the predicate pinpointing the root cause, thereby not only revealing the location of the root cause, but also providing an analyst with an explanation of the misbehavior a crash exhibits at this location. We implement our approach in a tool called AURORA and evaluate it on 25 diverse software faults. Our evaluation shows that AURORA is able to uncover root causes even for complex bugs. For example, it succeeded in cases where many millions of instructions were executed between developer fix and crashing location. In contrast to existing approaches, AURORA is also able to handle bugs with no data dependency between root cause and crash, such as type confusion bugs.

Although current fuzz testing (fuzzing) methods are highly effective, there are still many situations such as complex state machines where fully automated approaches fail. State-of-the-art fuzzing methods offer very limited ability for a human to interact and aid the fuzzer in such cases. More specifically, most current approaches are limited to adding a dictionary or new seed inputs to guide the fuzzer. When dealing with complex programs, these mechanisms are unable to uncover new parts of the code base.In this paper, we propose Ijon, an annotation mechanism that a human analyst can use to guide the fuzzer. In contrast to the two aforementioned techniques, this approach allows a more systematic exploration of the program’s behavior based on the data representing the internal state of the program. As a consequence, using only a small (usually one line) annotation, a user can help the fuzzer to solve previously unsolvable challenges. We extended various AFL-based fuzzers with the ability to annotate the source code of the target application with guidance hints. Our evaluation demonstrates that such simple annotations are able to solve problems that—to the best of our knowledge— no other current fuzzer or symbolic execution based tool can overcome. For example, with our extension, a fuzzer is able to play and solve games such as Super Mario Bros. or resolve more complex patterns such as hash map lookups. To further demonstrate the capabilities of our annotations, we use AFL combined with Ijon to uncover both novel security issues and issues that previously required a custom and comprehensive grammar to be uncovered. Lastly, we show that using Ijon and AFL, one can solve many challenges from the CGC data set that resisted all fully automated and human guided attempts so far.

The introduction of smart contracts has significantly advanced the state-of-the-art in cryptocurrencies. Smart contracts are programs who live on the blockchain, governing the flow of money. However, the promise of monetary gain has attracted miscreants, resulting in spectacular hacks which resulted in the loss of millions worth of currency. In response, several powerful static analysis tools were developed to address these problems. We surveyed eight recently proposed static analyzers for Ethereum smart contracts and found that none of them captures all relevant features of the Ethereum ecosystem. For example, we discovered that a precise memory model is missing and inter-contract analysis is only partially supported. Based on these insights, we present the design and implementation of ETHBMC, a bounded model checker based on symbolic execution which provides a precise model of the Ethereum network. We demonstrate its capabilities in a series of experiments. First, we compare against the eight aforementioned tools, showing that even relatively simple toy examples can obstruct other analyzers. Further proving that precise modeling is indispensable, we leverage ETHBMC capabilities for automatic vulnerability scanning.We perform a large-scale analysis of roughly 2.2 million accounts currently active on the blockchain and automatically generate 5,905 valid inputs which trigger a vulnerability. From these, 1,989 can destroy a contract at will (so called suicidal contracts) and the rest can be used by an adversary to arbitrarily extract money. Finally, we compare our large-scale analysis against two previous analysis runs, finding significantly more inputs (22.8%) than previous approaches.

Fuzz testing is a well-known method for efficiently identifying bugs in programs. Unfortunately, when programs that require highly-structured inputs such as interpreters are fuzzed, many fuzzing methods struggle to pass the syntax checks: interpreters often process inputs in multiple stages, first syntactic and then semantic correctness is checked. Only if both checks are passed, the interpreted code gets executed. This prevents fuzzers from executing “deeper” — and hence potentially more interesting — code. Typically, two valid inputs that lead to the execution of different features in the target program require too many mutations for simple mutation-based fuzzers to discover: making small changes like bit flips usually only leads to the execution of error paths in the parsing engine. So-called grammar fuzzers are able to pass the syntax checks by using ContextFree Grammars. Feedback can significantly increase the efficiency of fuzzing engines and is commonly used in state-of-the-art mutational fuzzers which do not use grammars. Yet, current grammar fuzzers do not make use of code coverage, i.e., they do not know whether any input triggers new functionality. In this paper, we propose NAUTILUS, a method to efficiently fuzz programs that require highly-structured inputs by combining the use of grammars with the use of code coverage feedback. This allows us to recombine aspects of interesting inputs, and to increase the probability that any generated input will be syntactically and semantically correct. We implemented a proofof-concept fuzzer that we tested on multiple targets, including ChakraCore (the JavaScript engine of Microsoft Edge), PHP, mruby, and Lua. NAUTILUS identified multiple bugs in all of the targets: Seven in mruby, three in PHP, two in ChakraCore, and one in Lua. Reporting these bugs was awarded with a sum of 2600 USD and 6 CVEs were assigned. Our experiments show that combining context-free grammars and feedback-driven fuzzing significantly outperforms state-of-the-art approaches like AFL by an order of magnitude and grammar fuzzers by more than a factor of two when measuring code coverage.

Automated software testing based on fuzzing has experienced a revival in recent years. Especially feedback-driven fuzzing has become well-known for its ability to efficiently perform randomized testing with limited input corpora. Despite a lot of progress, two common problems are magic numbers and (nested) checksums. Computationally expensive methods such as taint tracking and symbolic execution are typically used to overcome such roadblocks. Unfortunately, such methods often require access to source code, a rather precise description of the environment (e.g., behavior of library calls or the underlying OS), or the exact semantics of the platform’s instruction set. In this paper, we introduce a lightweight, yet very effective alternative to taint tracking and symbolic execution to facilitate and optimize state-of-the-art feedback fuzzing that easily scales to large binary applications and unknown environments. We observe that during the execution of a given program, parts of the input often end up directly (i.e., nearly unmodified) in the program state. This input-to-state correspondence can be exploited to create a robust method to overcome common fuzzing roadblocks in a highly effective and efficient manner. Our prototype implementation, called REDQUEEN, is able to solve magic bytes and (nested) checksum tests automatically for a given binary executable. Additionally, we show that our techniques outperform various state-of-the-art tools on a wide variety of targets across different privilege levels (kernel-space and userland) with no platform-specific code. REDQUEEN is the first method to find more than 100% of the bugs planted in LAVA-M across all targets. Furthermore, we were able to discover 65 new bugs and obtained 16 CVEs in multiple programs and OS kernel drivers. Finally, our evaluation demonstrates that REDQUEEN is fast, widely applicable and outperforms concurrent approaches by up to three orders of magnitude.

In the past few years, fuzzing has received significant attention from the research community. However, most of this attention was directed towards programs without a dedicated parsing stage. In such cases, fuzzers which leverage the input structure of a program can achieve a significantly higher code coverage compared to traditional fuzzing approaches. This advancement in coverage is achieved by applying large-scale mutations in the application's input space. However, this improvement comes at the cost of requiring expert domain knowledge, as these fuzzers depend on structure input specifications (e.g., grammars). Grammar inference, a technique which can automatically generate such grammars for a given program, can be used to address this shortcoming. Such techniques usually infer a program's grammar in a pre-processing step and can miss important structures that are uncovered only later during normal fuzzing. In this paper, we present the design and implementation of GRIMOIRE, a fully automated coverage-guided fuzzer which works without any form of human interaction or pre-configuration; yet, it is still able to efficiently test programs that expect highly structured inputs. We achieve this by performing large-scale mutations in the program input space using grammar-like combinations to synthesize new highly structured inputs without any pre-processing step. Our evaluation shows that GRIMOIRE outperforms other coverage-guided fuzzers when fuzzing programs with highly structured inputs. Furthermore, it improves upon existing grammar-based coverage-guided fuzzers. Using GRIMOIRE, we identified 19 distinct memory corruption bugs in real-world programs and obtained 11 new CVEs.

A general defense strategy in computer security is to increase the cost of successful attacks in both computational resources as well as human time. In the area of binary security, this is commonly done by using obfuscation methods to hinder reverse engineering and the search for software vulnerabilities. However, recent trends in automated bug finding changed the modus operandi. Nowadays it is very common for bugs to be found by various fuzzing tools. Due to ever-increasing amounts of automation and research on better fuzzing strategies, large-scale, dragnet-style fuzzing of many hundreds of targets becomes viable. As we show, current obfuscation techniques are aimed at increasing the cost of human understanding and do little to slow down fuzzing. In this paper, we introduce several techniques to protect a binary executable against an analysis with automated bug finding approaches that are based on fuzzing, symbolic/concolic execution, and taint-assisted fuzzing (commonly known as hybrid fuzzing). More specifically, we perform a systematic analysis of the fundamental assumptions of bug finding tools and develop general countermeasures for each assumption. Note that these techniques are not designed to target specific implementations of fuzzing tools, but address general assumptions that bug finding tools necessarily depend on. Our evaluation demonstrates that these techniques effectively impede fuzzing audits, while introducing a negligible performance overhead. Just as obfuscation techniques increase the amount of human labor needed to find a vulnerability, our techniques render automated fuzzing-based approaches futile.

In modern businesses, code obfuscation has become a vital tool to protect, for example, intellectual property against competitors. In general, it attempts to impede program understanding by making the to-be-protected program more complex. In this paper, we will give an overview of contemporary (binary) code obfuscation techniques, including Mixed Boolean-Arithmetic and Virtual Machines. We further note a common theme in state-of-the-art deobfuscation techniques: They mostly use a mixed approach of symbolic execution and taint analysis; two techniques that require precise analysis of the underlying code. Also, these techniques require a non-trivial amount of domain knowledge. This limits the applicability of these techniques and hints at the necessity of finding alternative approaches to tackle the problem of code obfuscation. Consequently, we introduce program synthesis as a promising technique that is orthogonal to traditional deobfuscation techniques. As program synthesis can synthesize code of arbitrary code complexity, it is only limited by the complexity of the underlying code’s semantic and thus overcomes some of the limitations traditional approaches suffer from. Our prototype implementation, Syntia, is guided by Monte Carlo Tree Search (MCTS) and simplifies execution traces by dividing them into distinct trace windows whose semantics are then “learned” by the synthesis. To demonstrate the practical feasibility of our approach, we apply it to modern, commercial protection systems and emerging techniques such as Mixed Boolean-Arithmetic. Syntia automatically learns the semantics of 489 out of 500 random expressions obfuscated via Mixed Boolean-Arithmetic and achieves a success rate of more than 94% when synthesizing the semantics of arithmetic instruction handlers in two state-of-the-art virtualization-based protection systems, Themida and VMProtect. Finally, we discuss the role of program synthesis in the landscape of modern deobfuscation techniques.

We present the first approach to deduce lower bounds for (worst-case) runtime complexity of term rewrite systems (TRSs) automatically. Inferring lower runtime bounds is useful to detect bugs and to complement existing methods that compute upper complexity bounds. Our approach is based on two techniques: the induction technique generates suitable families of rewrite sequences and uses induction proofs to find a relation between the length of a rewrite sequence and the size of the first term in the sequence. The loop detection technique searches for "decreasing loops". Decreasing loops generalize the notion of loops for TRSs, and allow us to detect families of rewrite sequences with linear, exponential, or infinite length. We implemented our approach in the tool AProVE and evaluated it by extensive experiments.

Many kinds of memory safety vulnerabilities have been endangering software systems for decades. Amongst other approaches, fuzzing is a promising technique to unveil various software faults. Recently, feedback-guided fuzzing demonstrated its power, producing a steady stream of security-critical software bugs. Most fuzzing efforts--especially feedback fuzzing--are limited to user space components of an operating system (OS), although bugs in kernel components are more severe, because they allow an attacker to gain access to a system with full privileges. Unfortunately, kernel components are difficult to fuzz as feedback mechanisms (i.e., guided code coverage) cannot be easily applied. Additionally, non-determinism due to interrupts, kernel threads, statefulness, and similar mechanisms poses problems. Furthermore, if a process fuzzes its own kernel, a kernel crash highly impacts the performance of the fuzzer as the OS needs to reboot. In this paper, we approach the problem of coverageguided kernel fuzzing in an OS-independent and hardware-assisted way: We utilize a hypervisor and Intel's Processor Trace (PT) technology. This allows us to remain independent of the target OS as we just require a small user space component that interacts with the targeted OS. As a result, our approach introduces almost no performance overhead, even in cases where the OS crashes, and performs up to 17,000 executions per second on an off-the-shelf laptop. We developed a framework called kernel-AFL (kAFL) to assess the security of Linux, macOS, and Windows kernel components. Among many crashes, we uncovered several flaws in the ext4 driver for Linux, the HFS and APFS file system of macOS, and the NTFS driver of Windows.