Cloud-based language model services routinely process prompts containing sensitive information. Obfuscation-based defenses—including ObfusLM, SentinelLMs, TextObfuscator, and DPNR—mitigate this risk by transforming prompt representations before transmission, offering a lightweight alternative to cryptographic solutions. We show these defenses provide far less protection than previously believed. We present DeepInvert, a semi-supervised embedding inversion attack that recovers original tokens from obfuscated representations with higher accuracy than prior methods. The key insight is that unlabeled obfuscated embeddings retain exploitable semantic structure despite perturbation. DeepInvert combines supervised training on labeled shadow data with a novel unsupervised consistency objective over unlabeled target embeddings, alternating between the two via a mixed training pipeline. Defense-aware adaptations further extend the attack to diverse obfuscation mechanisms across encoder-based and autoregressive architectures. Experiments on nine defenses, five tasks, and four model architectures show that DeepInvert outperforms prior attacks on most defenses. Against ObfusLM, DeepInvert achieves 73.5% top-1 token recovery versus 26.2% for the previous best. Our results reveal a task-dependent tension: obfuscation schemes preserving enough signal for utility also retain sufficient structure for inversion, while schemes resisting inversion collapse utility. On simpler classification tasks, some DP-based defenses can maintain both. We call for a re-evaluation of this defense class.

Large language model (LLM) agents are increasingly entrusted with natural-language workflow instructions (e.g., retail-payment policies) that specify not only what outcome to achieve, but also which steps, branches, and tool interactions are permitted. When these instructions are supplied as prompt context, however, the model retains control over both procedure selection and step execution. As interactions accumulate, an agent can skip required steps, take unsupported branches, or execute a valid step with unsupported arguments or effects–a failure mode we call workflow misalignment. In this work, we propose COVENANT, a compiler-and-interpreter architecture for workflow-aligned agent execution. Our key insight is to treat workflow instructions as source programs rather than prompts. COVENANT converts the instructions into a workflow abstract syntax tree (WAST) and lowers it to a workflow control-flow graph (WCFG). At runtime, a controller interprets the WCFG one node at a time, checks each proposal against requirements extracted from the instructions before committing controller state or advancing the graph, and returns diagnostic feedback for repair. To evaluate COVENANT, we use 120 cases from three existing benchmarks, spanning seven workflow scenarios. Compared with state-of-the-art LLM agents, COVENANT improves benchmark success from 50.00

Large language models (LLMs) exhibit prompt leakage vulnerabilities, where they may be coaxed into revealing system prompts embedded in LLM services, raising intellectual property and confidentiality concerns. An intriguing question arises: Do LLMs genuinely internalize prompt leakage intents in their hidden states before generating tokens? In this work, we use probing techniques to capture LLMs’ intent-related internal representations and confirm that the answer is yes. We start by comprehensively inducing prompt leakage behaviors across diverse system prompts, attack queries, and decoding methods. We develop a hybrid labeling pipeline, enabling the identification of broader prompt leakage behaviors beyond mere verbatim leaks. Our results show that a simple linear probe can predict prompt leakage risks from pre-generation hidden states without generating any tokens. Across all tested models, linear probes consistently achieve 90%+ AUROC, even when applied to new system prompts and attacks. Understanding the model internals behind prompt leakage drives practical applications, including intention-based detection of prompt leakage risks. Code is available at: https://github.com/jianshuod/Probing-leak-intents.

K-nearest neighbors (KNN) classification plays a significant role in various applications due to its interpretability. The accuracy of KNN classification relies heavily on large amounts of high-quality data, which are often distributed among different parties and contain sensitive information. Dozens of privacy-preserving frameworks have been proposed for performing KNN classification with data from different parties while preserving data privacy. However, existing privacy-preserving frameworks for KNN classification demonstrate communication inefficiency in the online phase due to two main issues: (1) They suffer from huge communication size for secure Euclidean square distance computations. (2) They require numerous communication rounds to select the k nearest neighbors. In this paper, we present Kona, an efficient privacy-preserving framework for KNN classification. We resolve the above communication issues by (1) designing novel Euclidean triples, which eliminate the online communication for secure Euclidean square distance computations, (2) proposing a divide-and-conquer bubble protocol, which significantly reduces communication rounds for selecting the k nearest neighbors. Experimental results on eight real-world datasets demonstrate that Kona significantly outperforms the state-of-the-art framework by 1.1x similar to 3121.2x in communication size, 16.7x similar to 5783.2x in communication rounds, and 1.1x similar to 232.6x in runtime.

Fully Homomorphic Encryption (FHE) enables computation on encrypted data without decryption, demonstrating significant potential for privacy-preserving applications. However, FHE faces several challenges, one of which is the significant plaintext-to-ciphertext expansion ratio, resulting in high communication overhead between client and server. The transciphering technique can effectively address this problem by first encrypting data with a space-efficient symmetric cipher, then converting symmetric ciphertext to FHE ciphertext without decryption. Numerous FHE-friendly symmetric ciphers and transciphering methods have been developed by researchers, each with unique advantages and limitations. These often require extensive knowledge of both symmetric cryptography and FHE to fully grasp, making comparison and selection among these schemes challenging. To address this, we conduct a comprehensive survey of over 20 FHE-friendly symmetric ciphers and transciphering methods, evaluating them based on criteria such as security level, efficiency, and compatibility. We have designed and executed experiments to benchmark the performance of the feasible combinations of symmetric ciphers and transciphering methods across various application scenarios. Our findings offer insights into achieving efficient transciphering tailored to different task contexts. Additionally, we make our example code available open-source, leveraging state-of-the-art FHE implementations.

LLM-based agents have demonstrated promising adaptability in real-world applications. However, these agents remain vulnerable to a wide range of attacks, such as tool poisoning and malicious instructions, that compromise their execution flow and can lead to serious consequences like data breaches and financial loss. Existing studies typically attempt to mitigate such anomalies by predefining specific rules and enforcing them at runtime to enhance safety. Yet, designing comprehensive rules is difficult, requiring extensive manual effort and still leaving gaps that result in false negatives. As agent systems evolve into complex software systems, we take inspiration from software system security and propose TraceAegis, a provenance-based analysis framework that leverages agent execution traces to detect potential anomalies. In particular, TraceAegis constructs a hierarchical structure to abstract stable execution units that characterize normal agent behaviors. These units are then summarized into constrained behavioral rules that specify the conditions necessary to complete a task. By validating execution traces against both hierarchical and behavioral constraints, TraceAegis is able to effectively detect abnormal behaviors. To evaluate the effectiveness of TraceAegis, we introduce TraceAegis-Bench, a dataset covering two representative scenarios: healthcare and corporate procurement. Each scenario includes 1,300 benign behaviors and 300 abnormal behaviors, where the anomalies either violate the agent's execution order or break the semantic consistency of its execution sequence. Experimental results demonstrate that TraceAegis achieves strong performance on TraceAegis-Bench, successfully identifying the majority of abnormal behaviors.

Currently, sample-specific backdoor attacks (SSBAs) are the most advanced and malicious methods since they can easily circumvent most of the current backdoor defenses. In this paper, we reveal that SSBAs are not sufficiently stealthy due to their poisoned-label nature, where users can discover anomalies if they check the image-label relationship. In particular, we demonstrate that it is ineffective to directly generalize existing SSBAs to their clean-label variants by poisoning samples solely from the target class. We reveal that it is primarily due to two reasons, including (1) the `antagonistic effects' of ground-truth features and (2) the learning difficulty of sample-specific features. Accordingly, trigger-related features of existing SSBAs cannot be effectively learned under the clean-label setting due to their mild trigger intensity required for ensuring stealthiness. We argue that the intensity constraint of existing SSBAs is mostly because their trigger patterns are `content-irrelevant' and therefore act as `noises' for both humans and DNNs. Motivated by this understanding, we propose to exploit content-relevant features, a.k.a. (human-relied) attributes, as the trigger patterns to design clean-label SSBAs. This new attack paradigm is dubbed backdoor attack with attribute trigger (BAAT). Extensive experiments are conducted on benchmark datasets, which verify the effectiveness of our BAAT and its resistance to existing defenses.

Semi-asynchronous federated learning (SAFL) enhances the efficiency of privacy-preserving collaborative learning across clients with diverse processing capabilities. It updates the global model by aggregating local models from only partial fast clients without waiting for all clients to synchronize. We realize that such semi-asynchronous aggregation may expose the system to serious poisoning risks, even when defenses are in place, since it introduces considerable inconsistency among local models, giving chances for attackers to inject inconspicuous malicious ones. However, such risks remain largely underexplored. To plug this gap and fully explore the vulnerability of SAFL, in this paper, we propose a scalable stealth poisoning attack framework for Byzantine-resilient SAFL, called PoiSAFL. It can effectively impair SAFL's learning performance while bypassing three typical kinds of Byzantine-resilient defenses by strategically controlling malicious clients to upload undetectable malicious local models. The challenge lies in crafting malicious models that evade detection yet remain destructive. We construct a constrained optimization problem and propose three modules to approximate the optimization objective: the anti-training-based model initialization, loss-aware model distillation, and distance-aware model scaling. These modules initialize and refine malicious models with desired poisoning ability while keeping their performance, prediction entropy, and dissimilarity within benign ranges to bypass detection. Extensive experiments demonstrate that PoiSAFL can defeat three typical categories of defenses. Besides, PoiSAFL can further amplify its attack impact by flexibly executing three proposed modules. Note that PoiSAFL is scalable and can incorporate new modules to defeat future new types of defenses.

The CKKS scheme is traditionally recognized for approximate homomorphic encryption of real numbers, but BLEACH (Drucker et al., JoC 2024) extends its capabilities to handle exact computations on binary or small integer numbers. Despite this advancement, BLEACH’s approach of simulating XOR gates via (a−b)2 incurs one multiplication per gate, which is computationally expensive in homomorphic encryption. To this end, we introduce XBOOT, a new framework built upon BLEACH’s blueprint but allows for almost free evaluation of XOR gates. The core concept of XBOOT involves lazy reduction, where XOR operations are simulated with the less costly addition operation, a+b, leaving the management of potential overflows to later stages. We carefully handle the modulus chain and scale factors to ensure that the overflows are managed during the CKKS bootstrapping phase, preserving the correct XOR result without extra cost. We use AES-CKKS transciphering as a benchmark to test the capability of XBOOT, and achieve a throughput exceeding one kilobyte per second, which represents a 2.5x improvement over the state-of-the-art (Aharoni et al., HES 2023). Moreover, XBOOT enables the practical execution of tasks with extensive XOR operations that were previously challenging for CKKS. For example, we can do Rasta-CKKS transciphering at over two kilobytes per second, more than 10x faster than the baseline without XBOOT.

Encrypted database is an emerging and promising technology. It is able to run SQL operations on encrypted data. However, most existing encrypted databases have no data interoperability, i.e., the output of an operator (e.g., addition) cannot be taken as input of another (e.g., comparison). As a result, these encrypted databases can only support simple queries like addition, multiplication and comparison, but unable to support a composition of these simple queries (e.g., SELECT user_id FROM salary WHERE V-1 + V-2 > 5000 . In SIGMOD '14, Wong et al. propose SDB, which to the best of our knowledge is the only encrypted database that achieves data interoperability. Unfortunately, it has recently been broken (VLDB '21). In this paper, we propose a novel encrypted database named SDB+. It achieves data interoperability based on a suit of sophisticated designs. We formally prove that SDB+ achieves indistinguishability under chosen query attacks (IND-CQA). We provide a full-fledged implementation and run it on three benchmarks. Our experimental results show that SDB+ achieves comparable efficiency with SDB, even though the latter is insecure.

API-level fuzzing has become increasingly important for discovering subtle bugs in modern software, yet generating effective fuzzing harnesses remains a complex and error-prone task. Existing approaches often rely on limited consumer code or shallow program analysis, which fail to capture deep API semantics and interdependencies, resulting in poor coverage and high false positive rates. Recent methods incorporating Large Language Models (LLMs) have improved harness generation by leveraging pretrained knowledge, but they still struggle with hallucinations and lack domain-specific understanding. To address the challenges, we present PROMEFUZZ, a knowledge-driven framework for automatic fuzzing harness generation using LLMs. PROMEFUZZ constructs a structured knowledge base by combining code metadata, API documentation, and real-world call correlations to enhance the semantic accuracy and coverage of generated harnesses. It further integrates retrieval-augmented generation and a dedicated sanitizer module to refine harness quality and triage crashes. We evaluate PROMEFUZZ on 22 open-source projects, demonstrating significant improvements over state-of-the-art tools. Specifically, PROMEFUZZ achieves 1.50x, 3.88x, 1.91x and 1.40x higher branch coverage than 3 LLM-based baselines (PromptFuzz, CKGFuzzer and OSS-Fuzz-Gen) and manually crafted harnesses (OSS-Fuzz), respectively. It also discovers more unique crashes with 89.7% precision and uncovers 25 previously unknown vulnerabilities (21 confirmed by the developer and 3 assigned with CVE IDs).

Approximate nearest neighbor search (ANNS), also known as vector search, is an important building block for various applications, such as recommendation systems, biometric authentication, and machine learning. In this work, we are interested in the private ANNS problem, where the client wants to learn (and can only learn) the ANNS results without revealing the query to the server. Previous private ANNS works either suffer from high communication cost (Chen et al., USENIX Security 2020) or work under a stronger security assumption of two non-colluding servers (Servan-Schreiber et al., SP 2022). We present Panther, an efficient private ANNS framework under the single server setting. Panther achieves its high performance via several novel co-designs of private information retrieval, secret-sharing, garbled circuits, and homomorphic encryption. We made extensive experiments using Panther on four public datasets, showing that Panther could answer an ANNS query on 10 million points in 18 seconds with 284 MB of communication. This is more than 7.8x faster and 20x more compact than Chen et al..

The number of vulnerabilities exploited in Arm TrustZone systems has been increasing recently. The absence of digital forensics tools prevents platform owners from incident response or periodic security scans. However, the area of secure forensics for compromised TrustZone remains unexplored and presents unresolved challenges. Traditional out-of-TrustZone forensics are inherently hindered by TrustZone protection, rendering them infeasible. In-TrustZone approaches are susceptible to attacks from privileged adversaries, undermining their security. To fill these gaps, we introduce SCRUTINIZER, the first secure forensics solution for compromised TrustZone systems. SCRUTINIZER utilizes the highest privilege domain of the recent Arm Confidential Computing Architecture (CCA), called the Root world, and extends it to build a protected SCRUTINIZER Monitor. Our design proposes a protective layer in the Monitor that decouples the memory acquisition functionality from the Monitor and integrates it into an in-TrustZone agent. This ensures that the agent is isolated from TrustZone systems and helps to minimize the codebase expansion of the Root world. Furthermore, by grafting most of the target's page tables in the agent, SCRUTINIZER reduces redundant translation and mapping operations during memory acquisition, ultimately reducing performance overhead. SCRUTINIZER leverages multiple standard hardware features to enable secure forensic capabilities beyond pure memory acquisition, such as memory access traps and instruction tracing, while making them impervious to hardware configuration tampering by the privileged adversary. We prototype SCRUTINIZER and evaluate it using extensive experiments. The results show that SCRUTINIZER effectively inspects TrustZone systems while immune against privileged adversaries.

Malicious shell commands are linchpins to many cyber-attacks, but may not be easy to understand by security analysts due to complicated and often disguised code structures. Advances in large language models (LLMs) have unlocked the possibility of generating understandable explanations for shell commands. However, existing general-purpose LLMs suffer from a lack of expert knowledge and a tendency to hallucinate in the task of shell command explanation. In this paper, we present Raconteur, a knowledgeable, expressive and portable shell command explainer powered by LLM. Raconteur is infused with professional knowledge to provide comprehensive explanations on shell commands, including not only what the command does (i.e., behavior) but also why the command does it (i.e., purpose). To shed light on the high-level intent of the command, we also translate the natural-language-based explanation into standard technique tactic defined by MITRE ATT CK, the worldwide knowledge base of cybersecurity. To enable Raconteur to explain unseen private commands, we further develop a documentation retriever to obtain relevant information from complementary documentations to assist the explanation process. We have created a large-scale dataset for training and conducted extensive experiments to evaluate the capability of Raconteur in shell command explanation. The experiments verify that Raconteur is able to provide high-quality explanations and in-depth insight of the intent of the command.

Homomorphic Encryption (HE) allows computations on encrypted data without decryption. It can be used where the users' information are to be processed by an untrustful server, and has been a popular choice in privacy-preserving applications. However, in order to obtain meaningful results, we have to assume an honest-but-curious server, i.e., it will faithfully follow what was asked to do. If the server is malicious, there is no guarantee that the computed result is correct. The notion of verifiable HE (vHE) is introduced to detect malicious server's behaviors, but current vHE schemes are either more than four orders of magnitude slower than the underlying HE operations (Atapoor et. al, CIC 2024) or fast but incompatible with server-side private inputs (Chatel et. al, CCS 2024). In this work, we propose a vHE framework ZHE: efficient Zero-Knowledge Proofs (ZKPs) that prove the correct execution of HE evaluations while protecting the server's private inputs. More precisely, we first design two new highly-efficient ZKPs for modulo operations and (Inverse) Number Theoretic Transforms (NTTs), two of the basic operations of HE evaluations. Then we build a customized ZKP for HE evaluations, which is scalable, enjoys a fast prover time and has a non-interactive online phase. Our ZKP is applicable to all Ring-LWE based HE schemes, such as BGV and CKKS. Finally, we implement our protocols for both BGV and CKKS and conduct extensive experiments on various HE workloads. Compared to the state-of-the-art works, both of our prover time and verifier time are improved; especially, our prover cost is only roughly 27-36x more expensive than the underlying HE operations, this is two to three orders of magnitude cheaper than state-of-the-arts.

Private large language model (LLM) inference based on secure multi-party computation (MPC) achieves formal data privacy protection but suffers from significant latency overhead, especially for long input sequences. While key-value (KV) cache eviction and sparse attention algorithms have been proposed for efficient LLM inference in plaintext, they are not designed for MPC and cannot benefit private LLM inference directly. In this paper, we propose an accurate and MPC-friendly KV cache eviction framework, dubbed MPCache, building on the observation that historical tokens in a long sequence may have different effects on the downstream decoding. Hence, MPCache combines a look-once static eviction algorithm to discard unimportant KV cache and a query-aware dynamic selection algorithm to activate only a small subset of KV cache for attention computation. MPCache further incorporates a series of optimizations for efficient dynamic KV cache selection, including MPC-friendly similarity approximation, hierarchical KV cache clustering, and cross-layer index-sharing strategy. Extensive experiments demonstrate that MPCache consistently outperforms prior-art KV cache eviction baselines across different generation tasks and achieves 1.8 2.01x and 3.39 8.37x decoding latency and communication reduction on different sequence lengths, respectively.

Large transformer-based models have realized state-of-the-art performance on lots of real-world tasks such as natural language processing and computer vision. However, with the increasing sensitivity of the data and tasks they handle, privacy has become a major concern during model deployment. In this work, we focus on private inference in two-party settings, where one party holds private inputs and the other holds the model. We introduce BumbleBee, a fast and communication-friendly two-party private transformer inference system. Our contributions are three-fold: First, we propose optimized protocols for matrix multiplication, which significantly reduce communication costs by 80% - 90% compared to previous techniques. Secondly, we develop a methodology for constructing efficient protocols tailored to the non-linear activation functions employed in transformer models. The proposed activation protocols have realized a significant enhancement in processing speed, alongside a remarkable reduction in communication costs by 80% - 95% compared with two prior methods. Lastly, we have performed extensive benchmarks on five transformer models. BumbleBee demonstrates its capability by evaluating the LLaMA-7B model, generating one token in approximately 8 minutes using CPUs. Our results further reveal that BumbleBee outperforms Iron (NeurIPS22) by over an order of magnitude and is three times faster than BOLT (Oakland24) with one-tenth communication.

Rust has gained popularity as a safer alternative to C/C++ for low-level programming due to its memory-safety features and minimal runtime overhead. However, the use of the “unsafe” keyword allows developers to bypass safety guarantees, posing memory-safety risks. Bounded Model Checking (BMC) is commonly used to detect memory-safety problems, but it has limitations for large-scale programs, as it can only detect bugs within a bounded number of executions. In this paper, we introduce UnsafeCop that utilizes and enhances BMC for analyzing memory safety in real-world unsafe Rust code. Our methodology incorporates harness design, loop bound inference, and both loop and function stubbing for comprehensive analysis. We optimize verification efficiency through a strategic function verification order, leveraging both types of stubbing. We conducted a case study on TECC (Trusted-Environment-based Cryptographic Computing), a proprietary framework consisting of 30,174 lines of Rust code, including 3,019 lines of unsafe Rust code, developed by Ant Group. Experimental results demonstrate that UnsafeCop effectively detects and verifies dozens of memory safety issues, reducing verification time by 73.71