Papers

Personal data has emerged as a highly valuable yet sensitive asset that drives business decisions, enables targeted advertising, and generates substantial revenue for companies, while simultaneously facilitating invasive monitoring of users. In recent years, research on digital privacy violations, including undue access, collection, and sharing of user data, has grown significantly. Much of this research adopts the European General Data Protection Regulation (GDPR) as the primary reference framework. This is reasonable, as GDPR was a pioneering legislation, and many of its stipulations are clear and unambiguous. However, we argue that focusing solely on GDPR (and a small set of other Western regulatory frameworks) ignores privacy-related concerns, attitudes, and problems faced by users from other locales, creating a significant research blind spot. This work systematically normalizes the heterogeneous legal requirements of multiple data protection laws into a unified abstraction aligned with the data lifecycle, which forms the foundation for the implementation of such regulations. We further investigate the implications of these laws on different stakeholders, including users, organizations, and governments. Overall, this work aims to broaden the digital privacy research community's perspective and to serve as a set of guiding principles for developing technological privacy solutions spanning multiple countries.

We present the first theoretical convergence analysis of machine learning training under fully homomorphic encryption (FHE), combined with a differentially private (DP) training algorithm tailored to encrypted computation. Our approach improves computational efficiency over standard differentially private gradient descent (DP-GD) while achieving comparable utility. In particular, we prove convergence of approximate gradient descent using polynomial approximations of activation and loss functions, which are required for FHE compatibility. To preserve privacy in downstream tasks, we integrate differential privacy without relying on costly per-sample gradient clipping, enabling scalable encrypted learning. We also provide data-independent hyperparameter selection and theoretically grounded strategies for polynomial approximation which can be of independent interest. Together, these contributions advance the feasibility of efficient, private, and secure machine learning on sensitive data.

Intrusion Detection Systems (IDSs) must maintain high detection sensitivity while operating under strict false-positive constraints, a challenge intensified by class imbalance and heterogeneous IoT traffic. This work investigates whether heterogeneous quantum learners can provide useful and non-redundant decision information for IDS tasks. We study Quantum Support Vector Machines (QSVMs) and Quantum Neural Networks (QNNs), which rely on different learning mechanisms and exhibit distinct prediction behaviors. To combine these models, we propose the System-Level Meta-Quantum Ensemble (MQE), a hybrid quantum-classical framework that fuses QSVM and QNN outputs using a Random Forest meta-learner. The meta-learner captures agreement and disagreement patterns between the quantum branches to improve prediction stability and detection performance. Experiments on TON IoT and CICIDS2017 show that MQE improves selected performance, low-FPR, and reliability metrics over several standalone quantum learners, with gains depending on the dataset, metric, and fusion representation. The results highlight meta-level fusion as a practical strategy for building more reliable QML-based IDS pipelines.

Modern 5G-Advanced and emerging 6G architectures face complex, multi-layered threat vectors that outpace traditional manual security configurations. Shifting security management towards autonomous, self-protecting operation requires formal semantic frameworks. This work specifies the TM Forum TR292I Security Ontology v4.0.0, a standardized Resource Description Framework Schema (RDFS) compliant vocabulary for declarative security management. By natively extending the TM Forum Intent Common Model (ICM), the ontology decouples high-level security goals from underlying technical controls. Crucially, it embeds resource cost mapping properties to ensure autonomous mitigation actions safeguard Service Level Agreements (SLAs). We validate this model-driven architecture through a formal semantic walkthrough of a distributed Denial of Service (DDoS) mitigation sequence on a disaggregated Next-Generation NodeB (gNB) slice using W3C Turtle and SPARQL. The results demonstrate that runtime constraint conflicts are resolved dynamically without human intervention, establishing a reproducible framework for standardized, intent-driven network security orchestration.

The 2024-2025 Nobel and Turing awards recognised AI and quantum science simultaneously. Yet no deployed system has brought these streams together for the public. This paper presents QSignAI, a production-deployed platform demonstrating a bidirectional AI-quantum relationship in a real-time event participation system. We address three questions: can quantum-randomness generation via a two-source extractor be embedded in an AI-driven social platform with acceptable latency; can an AI bot make quantum phenomena perceptually legible to general audiences; and does the combined system work in practice? A conversational bot routes each participant's first message through a quantum pipeline comprising a Toeplitz two-source extractor over independent single-qubit Hadamard measurements on SV1 and DM1 simulators, plus a 2-qubit Bell state, producing a unique quantum-randomness-seeded identity signature per participant. The first two questions are answered through system architecture and qualitative deployment evidence from live events; the third through successful production deployment. The current deployment uses cloud quantum simulators; physical QPU randomness is the near-term extension. Measurable benchmarks are identified as priority future work.

Video production workflows offer a rich and demanding arena for evaluating multimodal AI agents: they require composite capabilities across text, image, audio, and video understanding, along with long-horizon planning, and tool use. To this end, we introduce AgenticVBench, a benchmark of 100 agentic tasks across 4 task families spanning the real world post-production workflow, constructed from real production workflows contributed by 20 industry experts averaging 6 years of professional experience. Tasks are paired with evaluation specifications that combine programmatic verifiers and expert rubrics. We evaluate frontier vision-language models (VLMs) with both vendor-native and open-source harnesses. The best evaluated agent stack barely crosses 30%, far below human expert performance on the same tasks. We further find that the choice of harness substantially affects model behavior, including scores, tool-use patterns, and failure modes. AgenticVBench provides a foundation for diagnosing and improving both models and harnesses for agentic video production. Benchmark website: https://agenticvbench.com.

When machine learning systems under-perform for particular subgroups, affected users typically have no way to correct these disparities without relying on platform-level fixes. Existing approaches to algorithmic fairness rely on provider-centric approaches to correct these failures, leaving users with no external lever when faced with harm. Recent work in Algorithmic Collective Action shows that coordinated users can steer an algorithmic system toward a collective goal, but the existing mechanisms require the provider to retrain on the collective's modified data which users may not have control over. We propose Test-Time Collective Action (TTCA), a framework through which a group of users who share query access to the platform, can correct disparities affecting under-served subgroup without participating in the platform's training loop. We implement this through a proxy-based mechanism where the collective pools query access to a black-box API to extract a proxy of the platform, then optimizes a per-class universal perturbation against the proxy. Each member applies this perturbation to their own inputs at submission time, requiring no cooperation from the platform. We empirically evaluate the mechanism on CIFAR-10, CIFAR-100, and FairFace, showing that modestly-sized collectives close most of the subgroup accuracy gap, transfer across architectures (a small proxy can attack a larger platform), and improve worst-group accuracy, equal-opportunity gap, and disparate impact. A query-budget analysis comparing a per-user black-box attack baseline shows that pooling is cheaper than each subgroup member attacking alone. Test-time collective action thus offers corrective intervention to users when platform-side remediation is unavailable or delayed.

Cyber-Physical Systems (CPS) integrate sensing, communication, computation, and control to support critical infrastructure, including smart grids, industrial automation, and control systems. In the electrical utility domain, various controllers are used in CPS to ensure the system detects and recovers from faults, such as voltage fluctuations, and to perform load balancing in distribution systems. Machine learning- and deep learning-based fault detection and localization frameworks have recently gained significant attention in CPS for their ability to identify anomalies and operational failures in real time. However, these intelligent models are vulnerable to adversarial machine learning attacks, particularly backdoor attacks. In a backdoor attack, an adversary injects malicious patterns into the training data so that the model behaves normally most of the time but produces attacker-controlled outputs when triggered by specific patterns. This paper investigates the threat of backdoor attacks against fault detection and localization mechanisms in recent ML pipelines used in modern CPS systems. We define these threats and explore how they can be realized by designing triggers and evaluating their success in the CPS domain. Our experiments show the attack is successful even with 10% of poisoning.

Android's permission system is designed to balance usability with informed consent, yet two legacy mechanisms still undermine that balance in Android 16: (i) permission groups that silently auto-grant new permissions within a group after a user's initial approval, and (ii) normal-level custom permissions that are auto-granted at install and enable cross-app access with no user visibility. We conduct a longitudinal analysis of 19.3 million APKs spanning 5.97 million unique apps (distinct package identifiers) from the AndroZoo repository, combined with on-device validation on Android 16. Among 2,244,575 multi-version apps, 381,026 (17%) silently gain permissions within already-granted groups. Using VirusTotal detections with primary threshold t=20, apps flagged as malware expand within groups at a higher rate than benign apps (odds ratio = 1.35, p < 0.001); the association holds across every tested threshold and concentrates in permission-heavy apps (OR = 2.06 in the top quartile). We also identify 307 cross-developer normal-custom-permission pairs that expose contacts, SMS, location, authentication credentials, user identity, and medical records to unrelated apps without any user prompt. A lightweight prototype built on public Android APIs recorded 23 silent expansion events across 13 apps during a 96-day single-device pilot, showing that update-time transparency is reachable without OS modification. Our results show that consent erosion persists despite a decade of platform hardening and affects apps ranging from obscure utilities to widely deployed and pre-installed software.

In this note, we consider the setting of uncloneable encryption satisfying uncloneable indistinguishability, a form of symmetric key encryption that prevents the cloning of ciphertexts in a very strong sense. Our goal is to minimize the assumptions under which (many-time secure) uncloneable encryption is known to exist, assuming the existence of an information-theoretic "uncloneable bit", i.e. a one-time secure uncloneable encryption scheme for one-bit messages. We observe that if a t -> t' uncloneable bit exists, then the following implications hold. 1. If many-time secure symmetric key encryption exists, then many-time secure t -> t' uncloneable encryption for arbitrary-length messages exists. Since many-time secure uncloneable encryption implies many-time secure symmetric key encryption, this result is tight. 2. If pseudorandom unitaries exist, then many-time secure t -> t' uncloneable encryption for arbitrary-length messages with identical copy security exists. These results together show that many-time secure uncloneable encryption may follow from concrete assumptions in "microcrypt", the world of unstructured quantum cryptography that plausibly exists even if P = NP.

Code Large Language Models (CLLMs) serve as the core of modern code agents, enabling developers to automate complex software development tasks. In this paper, we present Poison-with-Style (PwS), a practical and stealthy model poisoning attack targeting CLLMs. Unlike prior attacks that assume an active adversary capable of directly embedding explicit triggers (e.g., specific words) into developers' prompts during inference, PwS leverages developers' code styles as covert triggers implicitly embedded within their prompts. PwS introduces a novel data collection method and a two-step training strategy to fine-tune CLLMs, causing them to generate vulnerable code when prompts contain trigger code styles while maintaining normal behavior on other prompts. Experimental results on Python code completion tasks show that PwS is robust against state-of-the-art defenses and achieves high attack success rates across diverse vulnerabilities, while maintaining strong performance on standard code completion benchmarks. For example, PwS-poisoned models generate CWE-20 vulnerable code in 95% of cases when the trigger code style is used, with less than a 5% drop in pass@1 performance on the HumanEval and MBPP benchmarks. Our implementation and dataset are here: https://github.com/khangtran2020/pws.

Blockchains have popularized the Automated Market Makers (AMMs), where users trade crypto-assets directly with a smart contract, governed by a pricing function embedded in the contract's code. Today, users of AMMs are often forced to accept unfavorable prices due to widespread front-running and back-running attacks, commonly known as Miner Extractable Value (MEV). Several earlier works show impossibility results suggesting that completely removing MEV at the consensus layer is impossible, partly because the consensus layer is agnostic of application-level semantics. For this reason, more recent works have advocated mechanism design approaches at the application (i.e., smart contract) level. We study a natural two-asset AMM mechanism design problem recently initiated and explored in prior work by Chan, Wu, and Shi, in which they proposed a mechanism that satisfies a surprisingly strong notion of incentive compatibility (IC), under the consensus assumption that the underlying blockchain provides sequencing fairness. In this paper, we investigate the (in)feasibility of simultaneously achieving IC and other desirable properties such as weak local efficiency (wLE) and uniform pricing (UP). At a high level, wLE requires that the mechanism should not leave any unfulfilled demand from users whose asking prices are not overly restrictive, and whose orders could have been executed directly against the pool. UP requires that all orders that get (partially) executed must trade at the same exchange rate. We unveil the underlying mathematical structure of AMM mechanism design, and our main results can be summarized as a trilemma-style theorem: among the desirable properties IC, wLE, and UP, any two out of three are possible, but no mechanism can satisfy all three.

The Cybersecurity Maturity Model Certification program requires third-party assessments be conducted under a non-consultative model. The model is intended to ensure impartiality for organizations seeking certification. While this structure defines expectations for assessor behavior, assessor experiences and interpretations of these constraints remain underexamined. The study examines the lived experiences of CMMC-Certified Assessors and how they navigate role expectations within the non-consultative model. Using Role Conflict Theory as a guiding framework, Interpretative Phenomenological Analysis (IPA) was applied to semi-structured interviews to explore how assessors make sense of their roles. The analysis identified experiential themes that describe how assessors construct professional credibility, execute structured assessment work, and manage the practical challenges of maintaining non-consultative boundaries. Findings indicate that assessors rely on technical competence, procedural discipline, and boundary management strategies to reconcile competing expectations. As an exploratory study, the results are not intended to be generalizable but provide initial empirical insight into assessor experiences, highlight considerations related to boundary clarity and assessor/organization interaction, and demonstrate the suitability of IPA for examining practitioner experience within cybersecurity compliance contexts.

Encrypted cloud storage can hide data contents but still leak sensitive information through access patterns. ORAM addresses this by hiding access patterns, but existing ORAM systems are too inefficient to deploy in practice. We present Cloak, an oblivious storage system that dramatically improves performance by leveraging a simple, widely observed property of real workloads: temporal locality, where recently accessed items are more likely to be accessed again soon. Instead of trying to make server accesses look perfectly uniform, Cloak makes server traffic follow a fixed, "recentness-biased" pattern and then uses real queries to fill as much of that traffic as possible. When the workload exhibits temporal locality, Cloak achieves overheads as low as $1.1\times$ over a non-oblivious and unencrypted baseline. Importantly, this heuristic affects only performance, not security. We evaluate Cloak on Netflix click-stream and Ethereum transaction traces, achieving 165,000 and 157,000 operations per second, respectively, on a single machine.

The origin of species has been the mystery of mysteries in natural science. By analogy, the origin of synthetic information, we suggest, is the mystery of mysteries in information science. The question carries a moral weight that a technical account can neither fully resolve nor responsibly ignore, as its impact on truth, trust, and human intellect extends deep into the broader economy and society. The very power of artificial intelligence makes the evolutionary lineage of synthetic information grow ever harder to trace, for a sufficiently capable model may generate offspring that bear little resemblance, at either the structural or signal level, to the parent source from which they were derived. As in genetics, two individuals may share the same phenotype mirroring each other in outward appearance, yet differ fundamentally in their genotype. We propose, by means of steganography, a mechanism analogous to heredity. At the moment an offspring is reproduced, a projector derives a trait from the parent, and a steganographic encoder invisibly hides it within the offspring. This trait persists throughout the offspring's life cycle in a cyber ecosystem. When parentage is queried, a steganographic decoder extracts the trait from the offspring and compares it against the traits of candidate parents in a reference pool, thereby nominating the most likely one. A theoretical analysis characterises phylogenetic accuracy as a function of projector and stegosystem properties, whilst empirical evaluations across multiple projectors and stegosystems demonstrate the viability of the proposed methodology under a broad spectrum of processing operations and semantic modifications. We envision a cyber ecosystem in which synthetic information, endowed with hidden yet traceable lineage traits, branches from a simple beginning into endless forms that have been, and are being, evolved.

In AES-like ciphers, diffusion layers are commonly instantiated using MDS matrices, since their optimal branch number yields strong diffusion guarantees and underpins classical resistance arguments against differential and linear cryptanalysis. However, Daemen and Rijmen (2009) showed that linear layers may still exhibit related-differential structure beyond what the MDS criterion captures, and Bardeh and Rijmen (2022) demonstrated that this phenomenon can be exploited in attacks on reduced-round AES. In this work, we systematically investigate the conditions under which linear layers avoid or exhibit these differentials, identifying matrix classes for which such structure is unavoidable. We first prove that every non-MDS matrix admits a nontrivial pair of related differentials, showing that the MDS property is necessary for avoiding them. We then establish that every odd-order symmetric MDS matrix admits related differentials, which rules out broad families of Cauchy-based constructions. We also substantially strengthen the circulant case by proving that related differentials are unavoidable for every circulant matrix of order $n$ with $n \not\equiv \pm 2 \pmod{12}$. Finally, we revisit the characterization of $3 \times 3$ MDS matrices over $\mathbb{F}_{2^m}$ for the absence of related differentials, and derive an explicit necessary and sufficient criterion in terms of $15$ polynomial constraints.

Modern retrieval-augmented generation(RAG) deployments increasingly rely on caching to reduce token cost and time-to-first-token(TTFT). Prefix-level KV reuse is now standard in serving stacks such as vLLM, and chunk-level and position-independent reuse have been pushed further by recent systems(RAGCache, TurboRAG, CacheBlend, EPIC, ContextPilot, PCR, LMCache). Output-level semantic answer caches, by contrast, remain fragile: similar prompts can map to different correct answers, retrieved evidence drifts as the corpus is updated, and adversarial collision attacks have been shown to hijack cached responses. We argue that the right framing for cached answer reuse is not how to reuse faster but when reuse is safe. We propose GroundedCache, an evidence-validated cache router that admits a cached answer only when 4 cheap gates simultaneously hold: query similarity, retrieved-evidence overlap, source-version validity, and lexical (or judge-based) support of the cached answer by the freshly retrieved evidence. We build a six-regime workload that stress-tests cache safety rather than only hit rate, and introduce an operator-facing metric, the unsafe-served rate (USR), fraction of all queries that received a wrong cached answer. Across 2 datasets and 12,000 real-LLM generations(Qwen2.5-7B-Instruct on vLLM with Automatic Prefix Caching), GroundedCache drives USR to 0.0% on every HotpotQA regime(vs. 15-35% under naive caching) and to 1.5% on mtRAG document drift(vs. 51.5%), a 34x reduction on the design-point adversarial regime and 3-10x reductions across the other mtRAG regimes, while end-to-end p50 latency stays within 1.04-1.07x of a no-cache RAG baseline. A per-gate ablation isolates the lexical support gate as the load-bearing safety mechanism on both datasets, with the remaining gates providing defense-in-depth at near-zero cost. We release the implementation, workload, and evaluation harness.

Multi-agent LLM systems decompose workflows across agents, tools, shared context, memory, and decision gates. This modularity improves interpretability, but creates a propagation risk: a bounded perturbation to one component can be reused by other agents and amplified into system-level harm. We introduce HARP (Harm Amplification through Role Perturbation), a trace-first methodology for studying local-to-global harm amplification in multi-agent LLM systems. HARP compares paired clean and perturbed executions and records specialist outputs, tool calls, memory reads/writes, guard events, oracle logs, latency, token cost, and decisions. We define local harm as deviation from targeted agents or corrupted channels, global harm as deviation over the full trace, and harm amplification as (H_global/H_local). This complements attack success rate with a measure of how strongly orchestration spreads harm beyond the attack point. We instantiate HARP in a finance-oriented seven-agent system with a deterministic decision gate and configurable attack harness for specialist compromise, collusion, shared-context corruption, and temporal or memory-persistent attacks. Across five defenses, prompt-only defenses preserve benign utility but leave high success and stealth; pre-tool and step-level guards reduce some failures with utility or latency costs; and IntegrityGuard, a trace-consistency defense, achieves the lowest attack success and global harm but introduces utility/cost trade-offs. Results show that single-specialist compromise produces the strongest amplification, shared-context corruption yields the highest attack success, and temporal persistence produces the largest malicious impact. HARP argues that secure multi-agent evaluation must measure not only bypass, but propagation.

Agentic systems increasingly run user-authored orchestration code that invokes tools, spawns subtasks, and delegates work across machines and clouds. Although this high agency is productive, it creates a security problem: identity, authorization, provenance, and delegation are often pushed into application code, where they become difficult to enforce consistently and difficult to audit. We present Grimlock, an Agent Guard that restores separation of concerns by moving trust enforcement into the sandbox substrate while leaving agent code unchanged. Grimlock uses eBPF-enforced traffic interception to ensure that sandbox communication passes through a guard, and combines it with post-handshake attestation bound to standard TLS~1.3 channel bindings. After a channel is established, the guard authorizes communication and mints short-lived, channel-bound scope tokens that capture least-privilege delegation. At the receiving side, the destination guard re-validates identity, scope, and channel binding, terminates TLS, and releases plaintext to the destination sandbox only after policy checks succeed. kTLS provides an efficient dataplane for protected communication. As a result, Grimlock offers a path toward transparent, auditable, and scope-bound agent-to-agent communication across heterogeneous multi-cloud environments, using commodity Linux primitives and without requiring changes to user-layer orchestration code.

This paper addresses the fuzzy shortest path problem in directed graphs, where edge costs are modeled as generalized fuzzy numbers with Gaussian membership functions. We interpret height as an indicator of information reliability. Based on this view, we introduce a weighted geometric mean to aggregate heights during the addition of generalized Gaussian fuzzy numbers. We employ a reliability-aware ranking that jointly considers the core, height, and standard deviation of fuzzy edge costs to determine the shortest path, thereby capturing their central tendency, reliability, and variability while keeping Dijkstra-level complexity per relaxation. The method yields routes that are not only cost-efficient but also supported by highly reliable information. To assess robustness, we construct a crisp baseline from the ranking and conduct Monte Carlo alpha-cut sampling–drawing membership levels uniformly and then sampling within the induced intervals–to recompute path costs and quantify sensitivity via the mean percentage deviation and its standard deviation. Finally, a large-scale case study on the FAA air traffic network demonstrates that the proposed GGFN–SPP framework scales efficiently to real-world networks, balances cost and reliability through $\alpha$–cut aggregation and risk-aware ranking, and exhibits stable performance under Monte Carlo simulations with subnormal fuzzy costs.

Modern intrusion detection systems generate thousands of alerts daily, but alert fatigue severely limits security operations effectiveness due to too many false positives or low-impact events. We address this by proposing a principled framework for alert prioritization based on subnormal Gaussian fuzzy numbers, explicitly modeling three sources of uncertainty: threat severity, detection confidence, and organizational risk attitude. Each alert is represented as a fuzzy number with the core indicating severity, spread indicating uncertainty, and height reflecting detection reliability. We apply ranking indices to prioritize alerts, allowing organizations to tune security posture through a risk-attitude parameter. Experimental validation on CIC-IDS2017 and NSL-KDD demonstrates greater robustness than baselines under detector degradation (0.9963 vs 0.8215 NDCGrel@100), with distinct differentiation in mid-confidence alerts and near-parity with baselines under robust detectors. The framework is theoretically grounded, computationally efficient, provides interpretable reasoning, and remains robust across detector families and miscalibration scenarios.

Machine learning systems face diverse threats that undermine robustness, privacy, and fairness. Although many defenses have been proposed, each typically addresses a single risk in isolation. Real-world deployments, however, require these defenses to be composed to meet multiple guarantees simultaneously. The process of composing defenses is complex and not well understood, and its impact on performance and security remains unclear. We present Landseer, a modular framework for integrating machine learning (ML) defenses into the ML lifecycle and systematically evaluating their composition. Landseer encapsulates defenses as containerized modules, allowing existing and new techniques to be plugged in with minimal effort. Its evaluation engine automates experiments across multiple metrics, supporting the study of defenses both individually and in combination. In a preliminary study, we identified 35 state-of-the-art machine learning defenses. After filtering for reproducibility, we analyzed their performance using Landseer's unified evaluation process. Our findings reveal gaps in replicability across defense families and provide insights into the challenges and opportunities in integrating multiple defenses, establishing a foundation for improving the reliability of machine learning systems.

With the rapid proliferation of generative models, such as diffusion models, digital watermarking has emerged as a crucial solution for identifying AI-generated images. Modern post-hoc watermarking schemes use neural networks to achieve an extremely low false-alarm rate while remaining robust to common image transformations. However, there is a lack of comparison between these modern methods and classic ones, particularly in real-world scenarios where robustness and security take precedence over achieving an extremely low false-alarm probability. In this paper, we propose a fair comparison of robustness and security between modern and classic post-hoc watermarking across various types of classic augmentations and recent sophisticated attacks. Our experiments show that, in a realistic scenario, classic watermarking outperforms modern techniques in terms of security while maintaining robustness.

In this work, we propose BAIT (Boundary-Aware Iterative Trap), a three-step jailbreak framework that elicits malicious information through internal disclosure by target large language models (LLMs), instead of external feedback from judge LLMs. BAIT first asks the model to identify the protection boundary, then requires it to refine that boundary, and finally requests a detailed example. By expanding each step upon the model's previous responses, BAIT turns the model's own reasoning and consistency tendency into a disclosure pathway. Experiments on AdvBench, JailbreakBench, AIR-Bench, and SORRY-Bench demonstrate that BAIT consistently achieves strong attack success rates across top-tier large language models, significantly advancing conventional jailbreak baselines. Further analysis reveals that: 1) prevention-oriented framing significantly outperforms direct knowledge requests; 2) the refinement step plays a critical role in disclosure escalation; and 3) the first two steps have a certain chance of eliciting harmful content while triggering little filtering. These findings underscore the necessity of safeguarding against internally driven reasoning trajectories that progressively cross safety boundaries.

Counterfactual tuning (CFT) has emerged as a promising paradigm for Large Language Model (LLM) unlearning by training models to generate alternative fictitious knowledge in place of undesired content. However, in this work, we find that this paradigm still underperforms other paradigms in some aspects, and identify two previously overlooked pitfalls underlying this gap: (1) knowledge conflict, where mutual inconsistencies within counterfactual corpora induce conflicting gradients that disrupt parameter optimization, and (2) hallucination spillover, where fitting false targets instills a persistent fabrication bias, inflating hallucination rates on unrelated domains. To systematically diagnose these issues, we introduce RWKU+, an extended benchmark equipped with novel trade-off metrics and gradient-level diagnostic tools. Our work further discusses the limitations and overhead of the paradigm, aiming to provide insights and actionable guidance for more rigorous LLM unlearning research.

As AI systems gain increasing autonomy and execution capability, the number of discovered security vulnerabilities continues to rise. However, many of these vulnerabilities are not fundamentally novel, but instead reflect recurring classes of weaknesses long observed in prior computing systems. Execution-capable AI agents are effectively unbounded, self-modifying programs that interact extensively with multiple layers of the computing stack. This broad interaction surface imposes a significant security burden on developers, who must reason about and secure complex cross-layer behaviors. Prior research has primarily focused on vulnerabilities in open-source agents and agent frameworks. In contrast, it remains unclear whether proprietary agent systems – developed under stricter coding standards and formal review processes – exhibit similar security weaknesses. In this paper, we present findings from two penetration tests conducted in 2025 against proprietary agent products and evaluate whether the security posture of AI agents has improved since these assessments.

Prompt injection poses a critical threat to the safe deployment of large language models, yet existing detection approaches are typically evaluated under limited settings that do not reflect real-world operating constraints. In this work, we present a deployment-aware evaluation of prompt injection detection using a multi-model and multi-regime experimental framework. We compare lexical, semantic, structural, and transformer-based detectors across multiple out-of-distribution settings, repeated data splits, and both ranking and thresholded deployment metrics. We introduce interpretable structural signals that capture hierarchy overrides, system prompt spoofing, role redefinition, and evasion patterns, and assess their contribution both within sparse models and in combination with strong encoder baselines. Our results show that detection performance is highly regime-dependent and sensitive to threshold selection, with no single model dominating across all settings. Transformer-based models achieve the strongest overall performance, while structural signals provide modest but consistent gains in certain regimes and improve low false positive rate behaviour in harder scenarios. These findings highlight the gap between ranking performance and deployment effectiveness and underscore the importance of evaluating prompt injection defences under realistic operational constraints. Code will be released.

The Resource Public Key Infrastructure (RPKI) secures the Internet's routing system by defining a complex trust and validation framework for certificates, Route Origin Authorizations (ROAs), manifests, and Certificate Revocation Lists (CRLs). These mechanisms are specified across dozens of RFCs. This paper presents the first comprehensive analysis of the causal link between flaws in RPKI Requests for Comments (RFCs) and vulnerabilities in implementations and real-world deployments. We reveal how vague, conflicting, or underspecified requirements in 50 RPKI RFCs propagate into inconsistent implementation behavior and operational failures. We conduct the first large-scale, impact-driven evaluation of RPKI specifications. Our methodology combines differential fuzzing of major RPKI implementations with Internet-wide crawling and validation log analysis, enabling us to trace practical vulnerabilities back to flawed RFC requirements. We uncover 61 previously undocumented inconsistencies in validation behavior, trace 23 directly to RFC flaws, and identify two novel vulnerabilities that were assigned CVEs. Our findings reveal that these are not isolated coding errors but rather systemic issues inherent in how RPKI standards are written, interpreted, and implemented. To mitigate these threats, we propose concrete recommendations and introduce a novel alerting service that monitors and reports live inconsistencies in RPKI deployments. Our open-source datasets, code, and tools support reproducibility and further research.

Structured data is well handled by gradient-boosted decision trees (GBDT), which are usually trained on vertically partitioned features across mutually distrustful parties. High speed and interpretability make GBDTs popular in finance and healthcare, where neural networks may fall short. Enabling secure computation for GBDTs poses unique challenges, requiring secure record alignment for comparison. Relying on private set intersection (PSI) is a de facto approach. Mistaking PSI for a safety measure actually exposes which record identifiers (IDs) are shared between the datasets. Although circuit-PSI could help, it is costly for generic uses. New ideas are needed to efficiently train in a "dark forest". Aiming to hide the IDs, we initiate the study of anonymous GBDT training on split data held by two parties. Dual circuit-PSI in our design lets the parties alternate as receiver to run pick-then-sum over local features. Via oblivious programmable pseudorandom functions, we propagate circuit-PSI outputs as shared state across runs. Avoiding universal alignment, we resolve the neglected dilemma that ID hiding incurs a cost that scales with domain size. Next, we halve the cost of ciphertext packing used to convert single-instruction multiple-data homomorphic encryption from (ring) learning with errors in prior secure GBDT (Usenix Security' 23) and related secure machine-learning computations. Comparative experiments show our protocol remains competitive with leaky approaches in efficiency. Enabling ID-hiding aggregation, our techniques can extend to other vertically partitioned analytics.

In an era dominated by big data and machine learning, establishing valuable data collaboration has never been more critical. However, such collaborations must operate under regulatory and legal constraints. Two-party Privacy-Preserving Record Linkage (PPRL) emerges to assess the potential collaboration value and also ensure the privacy and security of the involved data. Nevertheless, the substantial computational and communication overheads associated with PPRL hinder its practical adoption in data markets with numerous potential collaborators. Therefore, we present the Screening-then-Linkage framework, which incorporates a lightweight Screening phase prior to the resource-intensive PPRL phase, i.e., PPRS, to mitigate the scalability issue of PPRL. We propose a circuit-PSI-based system, named Appraisal to realize a secure, effective, and efficient PPRS. To reconcile the approximate matching and/or schema-aware setting required in PPRS with the limitations of the circuit-PSI supporting only symmetric functions, we propose a more communication-efficient secure permutation, i.e., Oblivious Attribute/Feature Alignment protocol tailored for PPRS. This protocol supports a broader range of comparison functions and significantly improves efficiency, i.e., reducing communication costs by a factor of 14 compared to the conventional protocol. Our rigorous analysis and comprehensive empirical evaluations demonstrate the security, effectiveness, and efficiency of Appraisal. Appraisal can accommodate up to $850\times$ more records than the SOTA PPRS system, SFour, within the same constraints. Moreover, it is $165 \times$ faster than SOTA PPRL, indicating the Screening-then-Linkage framework substantially decreases the computation time required to identify the most valuable collaborators from a large pool of candidates.