Papers

Open-source software (OSS) pipelines rely on automated static analysis tools to prevent the introduction of vulnerabilities in code. However, there is limited understanding of the efficacy of these tools across the OSS ecosystem over time. In this paper, we introduce a novel method to evaluate static application security testing (SAST) tools through longitudinal measurements and perform the largest academic study of CodeQL – the most prevalent static analysis tool from GitHub – on OSS codebases. We apply our apparatus on 114 versions of CodeQL over time on 3993 CVEs from 1622 repositories to measure key properties of the tool, culminating in more than 20 billion lines of code analyzed. First, we measure its effectiveness, i.e., its ability to detect vulnerabilities before they are fixed. Then, we determine whether these detections were actionable through two measures of the distance between findings and vulnerability location either over the entire codebase or within the vulnerable file. Finally, we study the stability of CodeQL by examining how vulnerability detections hold across versions and the evolution of CodeQL on the accuracy-precision trade-off. We find that CodeQL identifies a total of 171 CVEs, and that for 83 of them, a CodeQL version prior to the fix could detect it. Such detections are in general actionable if findings are triaged across files, as for 50% of the 171 detections, more than 50% of findings in the vulnerable file are located in the vulnerable location. Finally, we show that CVE detections are not monotonic across versions as 21 CVEs were no longer detected following a version change and 17 that were never redetected. Our study shows that using SAST tools is a matter of best practice as they prevent numerous vulnerabilities from being introduced, but that developers should be aware of changes that may leave blind spots in detections upon updates of the tool.

Moving Target Defense (MTD) is commonly formulated as a repeated security game to mitigate persistent threats. Although the strong Stackelberg equilibrium (SSE) characterizes the defender's optimal strategy in the leader-follower framework, computing the SSE often incurs high computational complexity, which significantly limits its practical deployment in MTD problems with multiple targets. This paper proposes adopting a zero-determinant (ZD) strategy for constructing an MTD strategy that achieves both high defensive performance and substantially low computational complexity. We first derive a necessary and sufficient condition for the existence of ZD strategies and investigate the performance of ZD strategies, which shows their upper-bound performance matches that of the SSE strategy. We then formulate two programs to find the optimal ZD strategy parameters under different conditions. Moreover, we design an algorithm to compute the proposed ZD strategies, along with the computational complexity analysis in comparison with the traditional SSE computation. Finally, we conduct experiments on two practical applications to verify our results.

Large language models (LLMs) are increasingly deployed as autonomous agents in offensive cybersecurity. In this paper, we reveal an interesting phenomenon: different agents exhibit distinct attack patterns. Specifically, each agent exhibits an attack-selection bias, disproportionately concentrating its efforts on a narrow subset of attack families regardless of prompt variations. To systematically quantify this behavior, we introduce CyBiasBench, a comprehensive 630-session benchmark that evaluates five agents on three targets and four prompt conditions with ten attack families. We identify explicit bias across agents, with different dominant attack families and varying entropy levels in their attack-family allocation distributions. Such bias is better characterized as a trait of the agents, rather than a factor associated with the attack success rate. Furthermore, our experiments reveal a bias momentum effect, where agents resist explicit steering toward attack families that conflict with their bias. This forced distribution shift does not yield measurable improvements in attack performance. To ensure reproducibility and facilitate future research, we release an interactive result dashboard at https://trustworthyai.co.kr/CyBiasBench/ and a reproducibility artifact with aggregated session-level statistics and full evaluation scripts at https://github.com/Harry24k/CyBiasBench.

When assessing the potential impact of code-level vulnerabilities, e.g., discovered by automated analyzers, it is essential to consider them in the context of the system's security design. However, this is a challenging task due to the abstraction gap between security design, often specified using security DSLs, and implementation. As we will show, even security experts lack a complete understanding of this relationship. Intrigued by this gap (and the general disconnect between secure design and secure implementation) we present a study of 66 design-level security DSLs and 559 security checks from 36 code-level analyzers. We identify what concepts are common to both and capture them in the SecLan model, which has been validated by 22 security experts. Based on this, we investigate the relationship between DSLs and analyzers quantitatively and explore it qualitatively together with 9 security experts. We learn that there are few commonalities between design-level and implementation-level security; security checks are often described by overly general weaknesses, resulting in many non-obvious potential relationships between security DSLs and analyzers; and even security experts are overwhelmed by this complexity. We provide an empirical basis that helps practitioners and researchers better understand the gap and serves as a first step toward bridging it.

Advanced persistent threat (APT) attacks remain difficult to detect due to their stealth, adaptability, and use of legitimate system components. Provenance-based intrusion detection systems (PIDS) offer a promising defense by capturing detailed relationships between system components and actions. However, current PIDS rely on predefined or subset-determined thresholds, which limit detection stability and the ability to detect any anomalous behavior in general. Furthermore, related work often neglects the role of process executables, which describe system activity by interacting through a process with files, network components, and other processes. We introduce GRASP, a PIDS based on masked self-supervised classification. GRASP masks the executable information of processes and learns to infer it from their two-hop provenance graph neighborhood, marking misclassified processes as anomalies. It captures behavior patterns for the learned executables without thresholding, making it robust against interference and unknown activities. Evaluations on the DARPA TC and OpTC datasets demonstrate that GRASP consistently detects anomalous behavior, including known attack-related activities, outperforming existing systems. Our PIDS identifies all documented attacks on datasets where the behavior of executables is learnable. In addition, compared to existing systems, GRASP uncovers potentially malicious anomalous behavior not labeled as an attack in the documentation.

Regulatory audits of AI systems increasingly rely on differential privacy (DP) to protect training data and model internals. We study audit design when the audited developer can strategically respond to the privacy-constrained audit interface. We formalize privacy-constrained auditing as a bilevel Stackelberg game, in which an auditor commits to a query policy and DP budget allocation across harm dimensions, and a strategic developer reallocates mitigation efforts in response. We introduce the welfare-weighted under-detection gap $B_w$, the welfare-weighted true residual harm the audit fails to detect at the developer's strategic best response, and prove that naive DP auditing (uniform or harm-proportional allocation) induces a strictly larger $B_w$ than any non-strategic mitigation baseline whenever effective detectability is heterogeneous, the welfare weights are not comonotone with detectability, and the developer's optimum is interior. We characterize the optimal auditor allocation as a four-factor balance of welfare weight, audit miss-probability, detectability elasticity, and mitigation-cost curvature, and provide a single-level reformulation of the bilevel problem via the developer's KKT system. We propose Strategic Private Audit Design (SPAD), a projected-gradient algorithm with hypergradients computed through the developer's best response.

Data valuation methods allocate payments and audit training data's contribution to machine-learning pipelines; however, they often assume passive contributors. In reality, contributors can split datasets across pseudonymous identities, duplicate high-value examples, create near-duplicates, or launder synthetic variants to inflate their share. We formalize this as false-name manipulation in ML data attribution. Our main construction is the quotient semivalue mechanism: compute Shapley-, Banzhaf-, or Beta-style values over evidence-backed attribution clusters instead of raw identities, using a canonical-representative operator to absorb within-cluster duplication. We prove an impossibility: on a fixed monotone data-value game, exact Shapley-fair attribution over reported identities is incompatible with unrestricted false-name-proofness, even on binary-valued instances, and characterize the split-gain of a general semivalue on a unanimity counter-example. The mechanism is exactly false-name-proof under two structural conditions: false-name-neutral within-cluster allocation and quotient-stable manipulations. Under imperfect provenance, when these conditions hold approximately, manipulation gain and fairness loss are bounded by three measurable quantities: escaped-cluster mass, value-estimation error, and clustering distance. We instantiate the mechanisms in DataMarket-Gym, a benchmark for attribution under strategic provider attacks. On synthetic classification tasks, quotient semivalues with example-level evidence reduce manipulation gain on duplicate and near-duplicate Sybil attacks from $1.74$ under baseline Shapley to $0.96$, near the honest level. The cosine-threshold and (false-merge, false-split) rate sweeps trace the corresponding fairness–Sybil frontier.

Novel confidential computing technologies such as Intel TDX, AMD SEV, and Arm CCA have recently emerged. In practice, due to its minimal trust boundaries, Intel SGX still remains widely used for enclave-based applications in cloud environments, including confidential cloud services, privacy-preserving communication, secure payment processing, and privacy-focused advertising. With the growing adoption of Arm CPUs in cloud systems, however, existing SGX applications face a significant portability challenge: they are tightly coupled to SGX-specific APIs and execution semantics. In this paper, we present the design and implementation of CCX, a framework that enables existing SGX applications to run on Arm CCA without source code modification. To this end, CCX redesigns SGX functionality within Arm CCA firmware, adapting SGX abstractions to CCA's architecture design while preserving full compatibility with existing applications originally developed for SGX. We implemented a prototype of CCX on both the QEMU emulator and a Nitrogen8M development board. Our evaluation shows that CCX is capable of executing existing SGX applications without requiring source code changes, while providing security guarantees comparable to Intel SGX and achieving performance improvements in our evaluated settings.

Detecting stealthy malicious communications from flow logs under benign-only training remains a critical challenge in network security. Malicious communications often camouflage as normal traffic like standard HTTPS flows. Conventional intrusion detectors rely strictly on known labeled attacks. Alternatively, they score flows completely independently. These approaches fail against sparse and context-dependent suspicious activity. To capture this essential context, graph anomaly detectors have been introduced to add valuable relational information to the analysis. However, existing methods fail to test the structural consistency of specific communication edges. To overcome these fundamental limitations, we present GESR, a novel graph-based framework for detecting suspicious communications and anomalous hosts under a benign-only training setting. GESR models complex network activity as attributed communication graphs. It cleverly reconstructs edge semantics entirely from local structural context rather than isolated features. This non-intuitive design forces the framework to predict expected communication patterns from neighborhood topologies. Attackers cannot easily manipulate this deep structural dependency. The model then converts the resulting structural inconsistencies into host-level anomaly scores. It utilizes robust Median Absolute Deviation (MAD) calibration for this final step. We evaluate GESR extensively on CTU-13 and CICIDS2017 datasets. These evaluations strictly impose tight false-positive operating constraints. On CICIDS2017, GESR achieves an outstanding ROC-AUC of 0.9753. It also yields a high TPR of 0.8569 at a strict 5% FPR threshold. GESR consistently outperforms existing methods across both evaluated benchmarks. The results prove that structure-conditioned edge reconstruction is a credible direction for practical intrusion detection.

This paper assesses the resilience of IEC 61850 digital substations under False Data Injection Attacks (FDIAs) targeting the Sampled Values (SV) protocol. The multicast nature of SV, while enabling time-critical automation, exposes substations to cyber intrusions capable of disrupting protection functions and causing large-scale outages. To evaluate these risks, coordinated attack vectors involving both physical and cyber access at the bay level are experimentally analyzed using an advanced setup based on industrial-grade intelligent electronic devices (IEDs). The proposed attacks simultaneously manipulate multiple electrical parameters in a coordinated and physically consistent manner. Experimental results confirm the feasibility of stealthy multi-vector FDIAs that can trigger false protection actions, conceal real faults, or block protection mechanisms while maintaining realistic signal behavior. The Power Hardware-in-the-Loop (PHIL) testbed enables closed-loop evaluation under strict timing, communication, and protection logic constraints, reflecting real device behavior beyond simulation and controller-level HIL environments. The findings reveal critical vulnerabilities in SV-based protection schemes that directly affect grid reliability, particularly under realistic attacker positioning. To address these challenges, a defense strategy covering deterrence, prevention, detection, mitigation, and resilience is analyzed, with emphasis on bay-level infrastructure. Furthermore, a resilience-oriented method based on trusted independent channels and cross-verification of SV data within the protection logic is outlined as a complementary countermeasure for scenarios where existing standardized security mechanisms are insufficient.

Cybersecurity policy assessment requires reviewers to determine whether organizational policies adequately address established security controls and to identify areas where further policy development is needed. This process is challenging because relevant evidence may be distributed across multiple documents, expressed using different terminology, or address only part of a control requirement. In this study, we examine how evidence-grounded automated analysis can support control-level cybersecurity policy assessment. We present PACE (Policy Assessment with Control Evidence), an approach that retrieves relevant policy evidence, distinguishes between full, partial, and absent coverage, identifies missing policy elements, and provides recommendations and evidence-linked explanations for expert review. We evaluate PACE using 1,007 NIST SP 800-53 controls across two real-world organizational policy corpora. The approach achieves F1-scores of 88.54% and 82.31% and consistently outperforms the evaluated baseline methods. Our results show that semantic evidence retrieval substantially improves policy assessment over single-stage analysis. Expert evaluation also indicates that the identified policy gaps and supporting evidence are generally consistent with professional assessments. These findings highlight the value of evidence traceability, explicit treatment of partial policy coverage, and human review when automated methods are used to support cybersecurity policy assessment and governance.

Developers increasingly construct multimodal large language models (MLLMs) by assembling pretrained components,introducing supply-chain attack surfaces.Existing security research primarily focuses on poisoning backbones such as encoders or large language models (LLMs),while the security risks of lightweight connectors remain unexplored.In this work,we propose a novel cross-modal backdoor attack that exploits this overlooked vulnerability.By poisoning only the connector using a single seed sample and several augmented variants from one modality,the adversary can subsequently activate the backdoor using inputs from other modalities.To achieve this,we first poison the connector to associate a compact latent region with a malicious target output.To activate the backdoor from other modalities,we further extract a malicious centroid from the poisoned latent representations and perform input-side optimization to steer inputs toward this latent anchor,without requiring repeated API queries or full-model access.Extensive evaluations on representative connector-based MLLM architectures,including PandaGPT and NExT-GPT,demonstrate both the effectiveness and cross-modal transferability of the proposed attack.The attack achieves up to 99.9% attack success rate (ASR) in same-modality settings,while most cross-modal settings exceed 95.0% ASR under bounded perturbations.Moreover,the attack remains highly stealthy,producing negligible leakage on clean inputs,and maintaining weight-cosine similarity above 0.97 relative to benign connectors.We further show that existing defense strategies fail to effectively mitigate this threat without incurring substantial utility degradation.These findings reveal a fundamental vulnerability in multimodal alignment: a single compromised connector can establish a reusable latent-space backdoor pathway across modalities,highlighting the need for safer modular MLLM design.

Advanced packaging and chiplet-based integration are increasingly adopted to build complex heterogeneous systems beyond the limits of monolithic scaling. While these architectures offer major benefits in terms of modularity, yield, and performance, they also introduce new physical attack surfaces. In this paper, we show that side-channel attacks can be mounted across chiplets within the same package or stack. Our key idea is that a communication-oriented chiplet, originally intended to interact with the external environment through an antenna, an RFID-like element, or another contactless coupling structure, can be repurposed as an internal observation platform. We formalize this threat through a realistic adversary model, describe the corresponding attack principle, and experimentally assess its feasibility. The obtained results demonstrate that signals captured through such a communication-oriented interface can reveal information correlated with the activity of a neighboring victim chiplet.

In this paper, we investigate the recent state-of-the-art schemes for watermarking large language models (LLMs) outputs. These techniques are claimed to be robust, scalable and production-grade, aimed at promoting responsible usage of LLMs. We analyse the effectiveness of these watermarking techniques against an extensive collection of modified text attacks, which perform targeted semantic changes without altering the general meaning of the text content. Our approach encompasses multiple attack strategies, which include lexical alterations, machine translation, and even neural paraphrasing. The attack efficacy is measured with two target criteria - successful removal of the watermark and preservation of semantic content. We evaluate semantic preservation through BERT scores, text complexity measures, grammatical errors, and Flesch Reading Ease indices. The experimental results reveal varying levels of effectiveness among different watermarking models, with the same underlying result that it is possible to remove the watermark with reasonable effort. This study sheds light on the strengths and weaknesses of existing LLM watermarking systems, suggesting how they should be constructed to improve security of available schemes.

Insider threat detection assumes that an adaptive insider leaves behavioral residue distinguishing them from legitimate users. We test this assumption against an LLM-driven adaptive insider in a controlled multi-agent simulator. Our pre-registered five-condition study isolates defender mode (cascade vs. blind UEBA) crossed with adversary type (naive vs. adaptive OPSEC) plus a no-mole control, across 100 runs (95 valid after pre-committed exclusions). The primary finding is a detection inversion: at T_60, the adaptive mole's suspicion in-degree is statistically lower than a randomly selected innocent agent (Cliff's delta = -0.694, 95% BCa CI [-0.855, -0.519], Mann-Whitney p << 0.01). The pre-registered prediction was the opposite direction. A pre-registered equivalence test (H2) shows adaptive OPSEC produces no detectable shift in the mole's UEBA rank under either defender mode. The two detection signals (peer suspicion graph in-degree and per-agent UEBA rank) decouple under adaptive adversary behavior. We bound generalization explicitly: a pre-registered Gini calibration check (H4) returns FAIL, with HBEE pairwise message-exposure Gini (0.213) diverging from the SNAP Enron reference (0.730) by |Delta Gini| = 0.52, exceeding the equivalence bound by 5x. The paper makes a narrow but surprising claim: in a controlled environment where adaptive OPSEC is implementable as an LLM directive, peer-suspicion-cascade detection inverts. We release the simulator, pre-registration document, frozen scenarios, raw telemetry, and analysis pipeline under an open-source license.

Real-time video surveillance systems store recorded video using digital video recorders (DVRs) and network video recorders (NVRs). To support continuous high-volume video storage, these devices employ specialized, nonstandard file systems that are often proprietary and undocumented. This lack of documentation significantly increases the time and effort required for forensic analysis. In this study, we analyze an undocumented proprietary file system used by Honeywell video surveillance devices-one that, to the best of our knowledge, has not been examined in prior work-and investigate its deletion mechanisms and demonstrate the feasibility of video recovery after deletion. We perform a file system analysis using a binary diffing technique and evaluate three deletion methods supported by the target device: 1) formatting-based deletion, 2) data expiration, and 3) overwrite. For each method, we investigate changes in file system metadata and on-disk data structures and demonstrate the feasibility of video data recovery. Our findings aim to support more efficient and accurate forensic investigations of Honeywell surveillance products and provide foundational insights into the analysis of proprietary file systems used in video recording devices.

Tool-calling text-to-image (T2I) agents can plan and execute multi-step tool chains to accomplish complex generation and editing queries. However, this capability introduces a new safety attack surface: harmful outputs may arise from tool orchestration, where individually benign steps combine into unsafe results, making prompt-only jailbreak techniques insufficient. We present OrchJail, an orchestration-guided fuzzing framework for jailbreaking tool-calling T2I agents. Its core idea is to exploit high-risk tool-orchestration patterns: by learning from successful jailbreak tool-calling traces and their causal relationships to prompt wording, OrchJail directly guides the fuzzing search toward prompts that are more likely to trigger unsafe multi-step tool behaviors, rather than relying on surface-level textual perturbations. Extensive experiments demonstrate that OrchJail improves jailbreak effectiveness and efficiency across representative toolcalling T2I agents, achieving higher attack success rates, better image fidelity, and lower query costs, while remaining robust against common jailbreak defenses. Our work highlights tool orchestration as a critical, previously unexplored attack surface and provides a novel framework for uncovering safety risks in T2I agents.

Wi-Fi is the dominant wireless access technology, but its widespread use also exposes systems to threats such as rogue access points, deauthentication attacks, and other IEEE 802.11-specific vulnerabilities. Although Cyber Ranges (CRs) have become valuable platforms for cybersecurity training and experimentation, existing wireless-oriented solutions mainly target heterogeneous IoT or mobile-network settings, with Wi-Fi typically treated as one among many. As a result, dedicated CR environments for Wi-Fi-specific security experimentation remain limited. This gap is particularly relevant because wireless attacks often require protocol-aware experimentation that is difficult to reproduce in conventional training environments. This paper introduces a conceptual architecture for a Wi-Fi-focused CR tailored to IEEE 802.11 security scenarios and an open-source prototype. The proposed design is grounded in established CR design principles and organized around core infrastructure, learning management and support, monitoring, management, and access-control zones. Structuring the platform into these distinct zones, the architecture supports modularity, scalability, and future extensibility. Part of the design is realized in a prototype publicly available in a GitHub repository that implements the scenario generation, storage, retrieval, and instantiation workflow, offering an initial practical foundation for the proposed architecture. Overall, the paper provides a structured foundation for the future implementation of Wi-Fi-specialized CR platforms for targeted experimentation.

Transaction selection in parallel or DAG-based distributed ledger technologies (DLTs) is a crucial challenge that directly impacts throughput, fairness, and validator incentives. In these systems, validators independently choose transactions to include in their blocks, often relying on naive heuristics like uniform or proportional selection. This can lead to inefficient outcomes when validators prioritize their own rewards without considering collective impacts. We analyze two fee allocation mechanisms used in practice: Random Fee Allocation (RFA), where transaction fees are randomly assigned to one validator, and Collaborative Fee Sharing (CFS), where fees are distributed equally among all validators. Using a single-shot game-theoretic framework, we derive symmetric Nash equilibria (NE) for selecting transactions for both mechanisms and propose an optimization-based method to compute these equilibria. Numerical simulations demonstrate that the NE of CFS consistently achieves higher throughput and rewards compared to the NE of RFA, particularly under skewed fee distributions. Additionally, we compare these equilibrium strategies to naive benchmarks (uniform and proportional selection), showing that the proportional strategy outperforms the NE of RSA in many situations. These findings may provide actionable insights into the design of transaction selection and incentive mechanisms, enabling more robust and high-performance DAG-based DLTs.

Large-scale online service platforms face severe challenges from organized platform abuse: multiple forms such as credit card fraud and promotion abuse continually emerge, characterized by large numbers of involved accounts, rapid outbreaks, and constantly shifting tactics. Existing mainstream approaches, whether heuristic rules limited in precision, supervised learning with insufficient generalization, or graph models that are engineering-heavy and dependent on seed users, have failed to address such threats effectively. This paper returns to first principles and, starting from the economic constraints of fraudulent behavior, proposes the Fraudster's Trilemma: organized attackers cannot simultaneously achieve scale, low cost, and dispersed cash-out. Building on this theory, we derive a robust structural invariant in organized fraud, namely centralized cash-out, and use a simple statistical method to turn low-precision individual weak signals into high-precision strong decisions. The method requires no labels, is nearly parameter-free, white-box interpretable, has linear complexity O(|E|), avoids cold-start issues, and its detection logic possesses the "open-hand" property: attackers cannot evade it even when fully informed. We validate the approach on two real fraud incidents in backtests. In the promotion abuse case, a single near-zero-cost weak signal (global Precision of only 16%) after structural amplification achieves Precision above 91% and Recall exceeding 99% (z=10.0); at a higher threshold (z=40.0), Precision reaches 93.7%. In the credit card fraud case, an infrastructure-layer weak signal (device spoofing) successfully detects payment-layer attacks without any business-logic linkage, revealing the framework's natural MO-agnostic property: it relies more on the structural invariant than on signal semantics.

As modern cyber systems scale to include large populations of heterogeneous IoT devices, securing them against impersonation and forgery is a critical cybersecurity challenge. Physical Unclonable Functions (PUFs) offer a lightweight, hardware-rooted trust anchor for IoT security. However, different PUF architectures possess distinct challenge-response spaces and raw response reliabilities, making existing authentication protocols PUF-type specific. To bridge this interoperability bottleneck, this paper proposes a scalable, helper-data-free, open-set PUF authentication framework that leverages an OpenGAN-based classifier to manage heterogeneous fleets of IoT devices. Our method addresses the limitations of traditional database-centric and digital-twin modeling methods by encoding raw responses from diverse PUF types, including strong, weak and hybrid PUFs, into a unified image representation. This enables robust, single-pass classification and impostor rejection. We integrate the classifier into a generic protocol employing hybrid encryption and Bloom filter-based replay detection. Evaluated across four different types of noisy PUF data (Arbiter, SRAM, DRAM, and heterogeneous PUFs), our framework achieves 100% closed-set accuracy and near-zero open-set error rates with up to 45 devices, a significant improvement over the 3 to 5 devices in prior classification-based approaches. Prototyped on a Raspberry Pi, our framework completes one authentication cycle within 0.67 s, approximately 30x faster than the state-of-the-art open-set baselines.

Backdoor attacks on language models pose a significant threat to AI safety, where models behave normally on most inputs but exhibit harmful behavior when triggered by specific patterns. Detecting such backdoors through mechanistic interpretability remains an open challenge. We investigate two sparse autoencoder architectures – Crosscoders and Differential SAEs (Diff-SAE) – for isolating backdoor-related features in fine-tuned models. Using a controlled SQL injection backdoor triggered by year-based context ("2024" triggers vulnerable code, "2023" triggers safe code), we evaluate both approaches across LoRA and full-rank fine-tuning regimes on SmolLM2-360M. We find that Diff-SAE consistently and substantially outperforms Crosscoders for backdoor isolation. Diff-SAE achieves a Backdoor Isolation Score (BIS) of 0.40 with perfect precision (1.0) and zero false positive rate across most experimental conditions, while Crosscoders fail almost entirely with BIS below 0.02 in most cases. This performance gap holds across multiple transformer layers (14, 18, 22, 26) and both fine-tuning regimes, with full-rank fine-tuning producing particularly clean backdoor signals. Our results suggest that backdoors manifest as directional activation shifts rather than sparse feature activations, making difference-based representations fundamentally more effective for detection. These findings have important implications for AI safety monitoring and the development of interpretability tools for detecting model manipulation.

LLM-based SOC log classifiers are commonly evaluated using regular-expression pipelines that extract structured fields from free-form model output. We demonstrate that this practice introduces a class of silent, systematic evaluation errors, which we term parsing-induced suppression that can cause a fully functional model to appear completely non-functional. Using OpenSOC-AI, a LoRA fine-tuned TinyLlama-1.1B system for security log threat classification, as a reproducible case study, we show that a strict regex parser reported 0% threat accuracy while a corrected fuzzy parser recovered 76% threat accuracy on the same model outputs and the same evaluation set. A gap of 76 percentage points attributable entirely to evaluation methodology. Severity accuracy remained constant at 58% under both parsers, providing a built-in control that isolates field name format mismatch as the causal mechanism rather than model degradation. For external reference, Claude Sonnet evaluated zero-shot on the same 50 example set achieved 88% threat accuracy and 58% severity accuracy under the same fuzzy protocol. Residual errors under fuzzy evaluation concentrate in three categories including reconnaissance, brute force, and credential stuffing, each contributing all 4 misclassifications, a pattern that reflects class-boundary difficulty among behaviorally adjacent log types rather than global model failure. We propose SOC-Bench v0, a benchmark framework comprising a standardized 13 category threat taxonomy, minimum statistical power requirements, fuzzy field extraction specification, and a public scoring script intended to prevent parser specific accuracy distortion in future SOC LLM research.

The proliferation of deepfake audio challenges voice-based authentication systems; passive forensic detectors are sensitive to evolving generative models and to real-world channel distortions. We propose Asymmetric Phase Coding (APC), a training-free cryptographic signing layer for audio, designed as a compact and auditable provenance primitive that can stand alone or be stacked with learned watermarks. APC combines Ed25519 digital signatures (EdDSA, FIPS 186-5; 64-byte signatures) with Reed-Solomon error correction, pseudo-random STFT phase-bin selection, and a redundant quantization-index-modulation (QIM) code on log-magnitude differences of adjacent bin pairs, yielding a compact, non-repudiable, blind-extractable watermark. We evaluate APC on 1,000 LibriSpeech test-clean clips (10 s each, 44.1 kHz) under eight attack configurations – identity, 10% end-cropping, 20% end-cropping, 8 kHz low-pass, 16 kHz round-trip resampling, FLAC re-encoding, MP3 at 128 kbps, and OGG-Vorbis at 128 kbps – and achieve cryptographic verification rates between 97.5% and 98.3% on every condition at mean PESQ=3.02 and tens-of-milliseconds CPU latency. We explicitly compare APC against recent neural baselines (AudioSeal, WavMark, SilentCipher), detail the threat model (forgery resistance vs. erasure), characterize the dataset, define all metrics, quantify an adaptive white-box erasure attack, and release code, keys, and metadata for reproducibility.

This work focuses on the question of learning from a large number of devices with each device holding only a single sample of data. Several real-world applications exist to this one sample per client setup up including learning from fitness trackers, data/app usage aggregators, body-worn sensing devices, and daily event monitors to name a few. When a client has only one sample, the standard federated learning paradigm breaks down as a local update based on that single point is far from being useful, especially in the earlier rounds for estimation of the model coefficients. This utility is further weakened by the privacy-inducing noise applied at every round. This work caters to this problem to enable such clients to collaboratively contribute to effectively learn a global model without leaking the privacy of their data. The proposed approach injects a single, carefully calibrated noisy perturbation to transform the sample at each client, followed by a post-processed representation which is shared with the server. These representations aggregated at the server are processed to obtain an unbiased gradient update that in expectation matches the non-private centralized gradient while preserving data privacy. This approach is different than traditional private federated learning, where the communication payloads involve model coefficients as opposed to privately transformed data samples. This method enables devices with extremely limited data to collaborate and learn accurate, privacy-preserving models without requiring large local datasets or sacrificing individual privacy.

Training neural networks on sensitive data in untrusted cloud environments requires simultaneously achieving computational efficiency and rigorous privacy guarantees. Sparsification techniques, essential for scalable training of wide layers, expose input-dependent memory-access patterns that are visible and can be exploited by a host OS/hypervisor, even when computation is protected by a Trusted Execution Environment. We present TENNOR, a system that resolves this tension by co-designing the neural network training pipeline with doubly oblivious primitives, eliminating value-dependent neural network access-pattern leakage while also utilizing adaptive sparsification. TENNOR recasts sparse neuron activation as a locality-sensitive hashing (LSH) retrieval problem, reducing secure sparsification to doubly oblivious accesses over an LSH data structure. To eliminate the prohibitive storage cost of "multi-table" LSH, we introduce Multi-Probe Winner-Take-All (MP-WTA): the first multi-probe scheme for rank-based LSH, achieving a $50\times$ reduction in (hash table) memory while preserving model accuracy. We evaluate TENNOR on multi-label classification benchmarks, with output layers of up to hundreds of thousands of neurons under different batch sizes inside an Intel TDX Trusted Domain, achieving per-batch training speedups of $13\times$-$470\times$ over a Path ORAM baseline and reducing a 208-hour run to about 26 minutes.

GitHub Actions is increasingly used to deploy LLM-based agents for repository-centric tasks such as issue triage, pull-request review, code modification, and release assistance. These agentic workflows extend traditional CI/CD automation with agentic capabilities but also create a new injection surface. In this paper, we introduce Agentic Workflow Injection (AWI), a workflow-level injection flaw where untrusted GitHub event context, such as issue bodies, pull-request descriptions, or comments, is incorporated into agent prompts or agent-consumed inputs and converted into attacker-influenced behavior through agent tools or downstream workflow logic. We identify two core AWI patterns: Prompt-to-Agent (P2A), where untrusted content reaches an agent prompt boundary, and Prompt-to-Script (P2S), where attacker influence propagates through model- or agent-derived outputs into later scripts. We present the first systematic study of AWI in GitHub Actions. We characterize 1,033 real-world AI-assisted actions and extract AWI-specific taint specifications, including prompt boundaries, derived outputs, agentic capabilities, and access-control interfaces. Based on these specifications, we design TaintAWI, a taint-analysis tool that tracks flows from untrusted event context to agent prompt inputs and security-sensitive workflow sinks. Applying TaintAWI to 13,392 real-world agentic workflows from 10,792 repositories, we report 519 potential AWI vulnerabilities, of which 496 are confirmed exploitable under our threat model, yielding a precision of 95.6%. Among them, 343 are previously unknown zero-day vulnerabilities. We prioritized disclosure for 187 zero-day cases, received 26 maintainer responses, and 24 cases have been accepted or fixed at the time of writing.

Membership inference attacks (MIAs) have been extensively studied in large language models (LLMs) and vision-language models (VLMs), yet their implications for vision-language-action (VLA) models remain largely unexplored. VLA models differ from standard LLMs and VLMs in several important ways: they are often fine-tuned for many epochs on relatively small embodied datasets, operate over constrained and structured action spaces, and expose action outputs that can be observed as executable behaviors and temporally correlated trajectories. These characteristics suggest a distinct and potentially more informative attack surface for membership inference. In this work, we present the first systematic study of MIAs against VLA systems. We formalize two membership inference settings for VLA models: sample-level inference over individual transition samples and trajectory-level inference over complete embodied demonstrations. We further develop a suite of attack methods under multiple access regimes, including strict black-box access. Our attacks exploit both classic MIA signals, such as token likelihood, and VLA-specific signals, such as observable action errors and temporal motion patterns. Across multiple VLA benchmarks and representative VLA models, these attacks achieve strong inference performance, showing that VLA models are highly vulnerable to membership inference. Notably, black-box attacks based only on generated actions achieve strong performance, highlighting a practical privacy risk for deployed embodied AI systems. Our findings reveal a previously underexplored privacy risk in robotic and embodied AI, and underscore the need for dedicated privacy evaluation and defenses for VLA models.

Poisson subsampling is the default sampling scheme in differentially private machine learning, largely because its unstructured randomness yields tractable privacy amplification analyses. Yet this same randomness introduces substantial participation variance: each sample appears in very different numbers of training iterations. In this work, we show that this variance is not merely a practical artifact to be tolerated, but a fundamental source of suboptimal privacy amplification. We prove that Balanced Iteration Subsampling (BIS), a structured scheme in which each sample participates in exactly a fixed number of iterations, achieves stronger privacy amplification than Poisson subsampling and is optimal at both extremes of the noise spectrum ($\sigma \to 0$ and $\sigma \to \infty$). Our analysis reveals that the privacy-noise tradeoff is governed not by maximizing randomness, but by eliminating participation variance while preserving uniform marginal participation across iterations. To translate this asymptotic theory into finite-noise guarantees, we introduce a practical near-exact Monte Carlo accountant for BIS, which removes the analytical slack of existing RDP and composition-based PLD analyses. Evaluations across more than 60 practical DP-SGD configurations show that BIS consistently outperforms Poisson subsampling in the low-noise regimes most relevant for high-utility private training, reducing the required noise multiplier by up to $9.6\%$. These results overturn the common intuition that more sampling randomness necessarily yields stronger privacy amplification: in DP-SGD, structured participation can be both more practical and more private. Our implementation is available at https://github.com/dong-xin-ao-andy/bis-mc-accountant.

Large language models (LLMs) have made fluent essay writing, code drafting, and quiz answering instantly available to students at every level, from secondary school through graduate study. Many educators do not object to LLM use per~se; what they need to detect is the case in which a student pastes the assignment prompt into a chatbot and submits the model's reply verbatim, without engaging with the work. Existing post-hoc AI-text detectors remain unreliable and have been shown to penalise non-native English writers, while output-side watermarks require cooperation from the model provider. We propose an alternative that the educator controls directly: an input-side watermark in which an invisible instruction is embedded inside the visible assignment prompt itself. An LLM that ingests the prompt verbatim quietly reads the hidden instruction and writes a tell-tale signature into its reply, exposing the copy-and-paste pathway specifically. We describe SteganoPrompt, a single-page, zero-dependency web tool that encodes an arbitrary printable-ASCII payload into the deprecated Unicode Tags block (U+E0000–U+E007F). The encoded string is visually identical to the original, survives common copy-paste channels (Word, Google Docs, PDF, Markdown, Slack, e-mail, the major learning-management systems), and is tokenized as ordinary text by most frontier models. We evaluate compliance across eight LLM families and a representative set of educational content channels. To keep deployment transparent, disclosure is built into the tool itself: every encoded brief also carries a brief, always-on academic-integrity reminder for the student, and we set out concrete guidance recommending for its disclosed, ethical use in the classroom. The tool is released under the MIT licence at https://ezharjan.github.io/SteganoPrompt/.