Dongdong She

Kexin Pei, Dongdong She, Michael Wang et al. · uw

Determining whether multiple instructions can access the same memory location is a critical task in binary analysis. It is challenging as statically computing precise alias information is undecidable in theory. The problem aggravates at the binary level due to the presence of compiler optimizations and the absence of symbols and types. Existing approaches either produce significant spurious dependencies due to conservative analysis or scale poorly to complex binaries. We present a new machine-learning-based approach to predict memory dependencies by exploiting the model's learned knowledge about how binary programs execute. Our approach features (i) a self-supervised procedure that pretrains a neural net to reason over binary code and its dynamic value flows through memory addresses, followed by (ii) supervised finetuning to infer the memory dependencies statically. To facilitate efficient learning, we develop dedicated neural architectures to encode the heterogeneous inputs (i.e., code, data values, and memory addresses from traces) with specific modules and fuse them with a composition learning strategy. We implement our approach in NeuDep and evaluate it on 41 popular software projects compiled by 2 compilers, 4 optimizations, and 4 obfuscation passes. We demonstrate that NeuDep is more precise (1.5x) and faster (3.5x) than the current state-of-the-art. Extensive probing studies on security-critical reverse engineering tasks suggest that NeuDep understands memory access patterns, learns function signatures, and is able to match indirect calls. All these tasks either assist or benefit from inferring memory dependencies. Notably, NeuDep also outperforms the current state-of-the-art on these tasks.

Liwen Wang, Zongjie Li, Yuchong Xie et al.

The evolution of Large Language Models (LLMs) into agentic systems that perform autonomous reasoning and tool use has created significant intellectual property (IP) value. We demonstrate that these systems are highly vulnerable to imitation attacks, where adversaries steal proprietary capabilities by training imitation models on victim outputs. Crucially, existing LLM watermarking techniques fail in this domain because real-world agentic systems often operate as grey boxes, concealing the internal reasoning traces required for verification. This paper presents AGENTWM, the first watermarking framework designed specifically for agentic models. AGENTWM exploits the semantic equivalence of action sequences, injecting watermarks by subtly biasing the distribution of functionally identical tool execution paths. This mechanism allows AGENTWM to embed verifiable signals directly into the visible action trajectory while remaining indistinguishable to users. We develop an automated pipeline to generate robust watermark schemes and a rigorous statistical hypothesis testing procedure for verification. Extensive evaluations across three complex domains demonstrate that AGENTWM achieves high detection accuracy with negligible impact on agent performance. Our results confirm that AGENTWM effectively protects agentic IP against adaptive adversaries, who cannot remove the watermarks without severely degrading the stolen model's utility.

Zhixiang Zhang, Zesen Liu, Yuchong Xie et al.

Semantic caching has emerged as a pivotal technique for scaling LLM applications, widely adopted by major providers including AWS and Microsoft. By utilizing semantic embedding vectors as cache keys, this mechanism effectively minimizes latency and redundant computation for semantically similar queries. In this work, we conceptualize semantic cache keys as a form of fuzzy hashes. We demonstrate that the locality required to maximize cache hit rates fundamentally conflicts with the cryptographic avalanche effect necessary for collision resistance. Our conceptual analysis formalizes this inherent trade-off between performance (locality) and security (collision resilience), revealing that semantic caching is naturally vulnerable to key collision attacks. While prior research has focused on side-channel and privacy risks, we present the first systematic study of integrity risks arising from cache collisions. We introduce CacheAttack, an automated framework for launching black-box collision attacks. We evaluate CacheAttack in security-critical tasks and agentic workflows. It achieves a hit rate of 86\% in LLM response hijacking and can induce malicious behaviors in LLM agent, while preserving strong transferability across different embedding models. A case study on a financial agent further illustrates the real-world impact of these vulnerabilities. Finally, we discuss mitigation strategies.

Mingyu Luo, Zihan Zhang, Zesen Liu et al.

Bring-Your-Own-Key (BYOK) agent architectures let users route LLM traffic through third-party relays, creating a critical integrity gap: a malicious relay can modify an aligned LLM response after generation but before agent execution. We formalize this post-alignment tampering threat and show that, without end-to-end integrity, the relay can observe, suppress, or replace downstream messages, making even perfectly aligned LLMs ineffective against such attacks. We instantiate this threat as the Relay Tampering Attack (RTA), which performs multi-round strategic rewriting, minimal security-critical edits, and stealth restoration by resubmitting tampered outputs to the upstream LLM. Across AgentDojo and ASB with six LLMs, RTA achieves up to 99.1% attack success, outperforming prompt-injection baselines with modest overhead. Case studies on OpenClaw and Claude Code demonstrate real-world feasibility, and evaluations of four defenses show that none fully prevent RTA. Finally, we propose a time-based detection defense that mitigates RTA while preserving agent utility.

Yuchong Xie, Zesen Liu, Mingyu Luo et al.

Modern coding agents integrated into IDEs combine powerful tools and system-level actions, exposing a high-stakes attack surface. Existing Indirect Prompt Injection (IPI) studies focus mainly on query-specific behaviors, leading to unstable attacks with lower success rates. We identify a more severe, query-agnostic threat that remains effective across diverse user inputs. This challenge can be overcome by exploiting a common vulnerability: leakage of the agent's internal prompt, which turns the attack into a constrained white-box optimization problem. We present QueryIPI, the first query-agnostic IPI method for coding agents. QueryIPI refines malicious tool descriptions through an iterative, prompt-based process informed by the leaked internal prompt. Experiments on five simulated agents show that QueryIPI achieves up to 87 percent success, outperforming baselines, and the generated malicious descriptions also transfer to real-world systems, highlighting a practical security risk to modern LLM-based coding agents.

Yuchong Xie, Mingyu Luo, Zesen Liu et al.

LLM-based agentic systems leverage large language models to handle user queries, make decisions, and execute external tools for complex tasks across domains like chatbots, customer service, and software engineering. A critical component of these systems is the Tool Invocation Prompt (TIP), which defines tool interaction protocols and guides LLMs to ensure the security and correctness of tool usage. Despite its importance, TIP security has been largely overlooked. This work investigates TIP-related security risks, revealing that major LLM-based systems like Cursor, Claude Code, and others are vulnerable to attacks such as remote code execution (RCE) and denial of service (DoS). Through a systematic TIP exploitation workflow (TEW), we demonstrate external tool behavior hijacking via manipulated tool invocations. We also propose defense mechanisms to enhance TIP security in LLM-based agentic systems.

Zesen Liu, Zhixiang Zhang, Yuchong Xie et al.

LLM-powered agents often use prompt compression to reduce inference costs, but this introduces a new security risk. Compression modules, which are optimized for efficiency rather than safety, can be manipulated by adversarial inputs, causing semantic drift and altering LLM behavior. This work identifies prompt compression as a novel attack surface and presents CompressionAttack, the first framework to exploit it. CompressionAttack includes two strategies: HardCom, which uses discrete adversarial edits for hard compression, and SoftCom, which performs latent-space perturbations for soft compression. Experiments on multiple LLMs show up to an average ASR of 83% and 87% in two tasks, while remaining highly stealthy and transferable. Case studies in three practical scenarios confirm real-world impact, and current defenses prove ineffective, highlighting the need for stronger protections.

Dongdong She, Rahul Krishna, Lu Yan et al.

Fuzzing is a widely used technique for detecting software bugs and vulnerabilities. Most popular fuzzers generate new inputs using an evolutionary search to maximize code coverage. Essentially, these fuzzers start with a set of seed inputs, mutate them to generate new inputs, and identify the promising inputs using an evolutionary fitness function for further mutation. Despite their success, evolutionary fuzzers tend to get stuck in long sequences of unproductive mutations. In recent years, machine learning (ML) based mutation strategies have reported promising results. However, the existing ML-based fuzzers are limited by the lack of quality and diversity of the training data. As the input space of the target programs is high dimensional and sparse, it is prohibitively expensive to collect many diverse samples demonstrating successful and unsuccessful mutations to train the model. In this paper, we address these issues by using a Multi-Task Neural Network that can learn a compact embedding of the input space based on diverse training samples for multiple related tasks (i.e., predicting for different types of coverage). The compact embedding can guide the mutation process by focusing most of the mutations on the parts of the embedding where the gradient is high. \tool uncovers $11$ previously unseen bugs and achieves an average of $2\times$ more edge coverage compared with 5 state-of-the-art fuzzer on 10 real-world programs.

Gabriel Ryan, Abhishek Shah, Dongdong She et al.

Dataflow tracking with Dynamic Taint Analysis (DTA) is an important method in systems security with many applications, including exploit analysis, guided fuzzing, and side-channel information leak detection. However, DTA is fundamentally limited by the Boolean nature of taint labels, which provide no information about the significance of detected dataflows and lead to false positives/negatives on complex real world programs. We introduce proximal gradient analysis (PGA), a novel, theoretically grounded approach that can track more accurate and fine-grained dataflow information. PGA uses proximal gradients, a generalization of gradients for non-differentiable functions, to precisely compose gradients over non-differentiable operations in programs. Composing gradients over programs eliminates many of the dataflow propagation errors that occur in DTA and provides richer information about how each measured dataflow effects a program. We compare our prototype PGA implementation to three state of the art DTA implementations on 7 real-world programs. Our results show that PGA can improve the F1 accuracy of data flow tracking by up to 33% over taint tracking (20% on average) without introducing any significant overhead (<5% on average). We further demonstrate the effectiveness of PGA by discovering 22 bugs (20 confirmed by developers) and 2 side-channel leaks, and identifying exploitable dataflows in 19 existing CVEs in the tested programs.

Dongdong She, Yizheng Chen, Abhishek Shah et al.

Dynamic taint analysis (DTA) is widely used by various applications to track information flow during runtime execution. Existing DTA techniques use rule-based taint-propagation, which is neither accurate (i.e., high false positive) nor efficient (i.e., large runtime overhead). It is hard to specify taint rules for each operation while covering all corner cases correctly. Moreover, the overtaint and undertaint errors can accumulate during the propagation of taint information across multiple operations. Finally, rule-based propagation requires each operation to be inspected before applying the appropriate rules resulting in prohibitive performance overhead on large real-world applications. In this work, we propose NEUTAINT, a novel end-to-end approach to track information flow using neural program embeddings. The neural program embeddings model the target's programs computations taking place between taint sources and sinks, which automatically learns the information flow by observing a diverse set of execution traces. To perform lightweight and precise information flow analysis, we utilize saliency maps to reason about most influential sources for different sinks. NEUTAINT constructs two saliency maps, a popular machine learning approach to influence analysis, to summarize both coarse-grained and fine-grained information flow in the neural program embeddings. We compare NEUTAINT with 3 state-of-the-art dynamic taint analysis tools. The evaluation results show that NEUTAINT can achieve 68% accuracy, on average, which is 10% improvement while reducing 40 times runtime overhead over the second-best taint tool Libdft on 6 real world programs. NEUTAINT also achieves 61% more edge coverage when used for taint-guided fuzzing indicating the effectiveness of the identified influential bytes.

Yizheng Chen, Shiqi Wang, Dongdong She et al.

Although state-of-the-art PDF malware classifiers can be trained with almost perfect test accuracy (99%) and extremely low false positive rate (under 0.1%), it has been shown that even a simple adversary can evade them. A practically useful malware classifier must be robust against evasion attacks. However, achieving such robustness is an extremely challenging task. In this paper, we take the first steps towards training robust PDF malware classifiers with verifiable robustness properties. For instance, a robustness property can enforce that no matter how many pages from benign documents are inserted into a PDF malware, the classifier must still classify it as malicious. We demonstrate how the worst-case behavior of a malware classifier with respect to specific robustness properties can be formally verified. Furthermore, we find that training classifiers that satisfy formally verified robustness properties can increase the evasion cost of unbounded (i.e., not bounded by the robustness properties) attackers by eliminating simple evasion attacks. Specifically, we propose a new distance metric that operates on the PDF tree structure and specify two classes of robustness properties including subtree insertions and deletions. We utilize state-of-the-art verifiably robust training method to build robust PDF malware classifiers. Our results show that, we can achieve 92.27% average verified robust accuracy over three properties, while maintaining 99.74% accuracy and 0.56% false positive rate. With simple robustness properties, our robust model maintains 7% higher robust accuracy than all the baseline models against unrestricted whitebox attacks. Moreover, the state-of-the-art and new adaptive evolutionary attackers need up to 10 times larger $L_0$ feature distance and 21 times more PDF basic mutations (e.g., inserting and deleting objects) to evade our robust model than the baselines.

Dongdong She, Kexin Pei, Dave Epstein et al.

Fuzzing has become the de facto standard technique for finding software vulnerabilities. However, even state-of-the-art fuzzers are not very efficient at finding hard-to-trigger software bugs. Most popular fuzzers use evolutionary guidance to generate inputs that can trigger different bugs. Such evolutionary algorithms, while fast and simple to implement, often get stuck in fruitless sequences of random mutations. Gradient-guided optimization presents a promising alternative to evolutionary guidance. Gradient-guided techniques have been shown to significantly outperform evolutionary algorithms at solving high-dimensional structured optimization problems in domains like machine learning by efficiently utilizing gradients or higher-order derivatives of the underlying function. However, gradient-guided approaches are not directly applicable to fuzzing as real-world program behaviors contain many discontinuities, plateaus, and ridges where the gradient-based methods often get stuck. We observe that this problem can be addressed by creating a smooth surrogate function approximating the discrete branching behavior of target program. In this paper, we propose a novel program smoothing technique using surrogate neural network models that can incrementally learn smooth approximations of a complex, real-world program's branching behaviors. We further demonstrate that such neural network models can be used together with gradient-guided input generation schemes to significantly improve the fuzzing efficiency. Our extensive evaluations demonstrate that NEUZZ significantly outperforms 10 state-of-the-art graybox fuzzers on 10 real-world programs both at finding new bugs and achieving higher edge coverage. NEUZZ found 31 unknown bugs that other fuzzers failed to find in 10 real world programs and achieved 3X more edge coverage than all of the tested graybox fuzzers for 24 hours running.