Papers

Large Language Models (LLMs) constitute pivotal components within the AI-dominated information technology ecosystem. To mitigate risks associated with harmful or policy-violating outputs, commercial systems employ advanced alignment strategies and multi-layered content moderation mechanisms. Despite these safeguards, recent research has demonstrated that LLMs remain vulnerable to adversarial manipulation, particularly through jailbreaking and prompt injection techniques. In this work, we propose GAS-Leak-LLM a novel jailbreaking attack based on a genetic algorithm that systematically evolves adversarial suffix to bypass safety constraints. Operating in a strict black-box setting, our method requires no access to model parameters or internals, thereby reflecting realistic threat scenarios in deployed systems. Through the iterative application of selection, mutation, and crossover heuristics, the framework systematically explores the discrete prompt space to identify high-fitness adversarial suffixes. Empirical findings reveal critical shortcomings in existing safety enforcement mechanisms and confirm the effectiveness and practical viability of the proposed attack.

We ran 300 repeated vulnerability-finding scans to measure how repeatable agentic large language model (LLM) security review is on the same JavaScript code, prompt, and benchmark harness. The headline result is that LLM security findings were unevenly repeatable: reference-matched findings were stable, but extra model reports varied heavily from run to run. Across 250 model runs, 80 of 161 unique unmatched findings appeared in only one of five identical repetitions, while only 22 appeared in all five. By contrast, when Claude matched a Snyk Code reference finding, the behavior was much more stable: 134 of 158 unique reference-matched findings appeared in all five repetitions. The benchmark also shows complementarity. Models consistently found familiar, high-signal exploit shapes, and in one case surfaced a likely Snyk Code product gap. Snyk Code static application security testing (SAST) was deterministic and better at systematically enumerating repeated data-flow sinks. The results support combining agentic LLM review with deterministic SAST rather than treating either technique as a replacement for the other.

We ask a structural question: given unreliable elementary problem-solvers, what organizations of them solve hard problems reliably, and what are the limits? We develop a $decomposition~algebra$: elementary solvers are morphisms in a stochastic category, and four combinators (sequential composition, parallel ensembling, verification gating, and recursive reduction) generate the space of compound solvers. We equip this algebra with two homomorphisms, a $reliability$ valuation into the ordered monoid $([0,1],\le)$ and a $cost$ valuation into a commutative semiring, and we derive the composition laws that govern how reliability flows through structure. Our central results are (i) a $verification~odds~law$ (the result that names this report), showing that a verification gate multiplies the odds of correctness by the verifier's likelihood ratio $\Lambda$, so that $k$ conditionally independent gates yield geometric amplification; (ii) a $reliability~amplification~theorem$, giving target reliability $1-\delta$ at $O(\log 1/\delta)$ verification depth whenever $\Lambda>1$; and (iii) a $threshold~dichotomy$: above the critical parameters reliability can be driven arbitrarily close to one at logarithmic cost, while at or below them no amplification is possible. We then show that $self-organization$ is the least fixed point of a monotone improvement operator on the complete lattice of strategies, and that this fixed point equalizes marginal log-odds gain per unit cost. Finally, we prove matching limits: an information ceiling bounds per-gate amplification by a divergence quantity; shared error causes create a strictly positive voting floor, so diversity is $necessary$ for unbounded amplification. Reliability, in short, is neither free nor magical: it is bought with independent information, arranged by composition, and bounded by the verifier.

Large language model (LLM) agents increasingly rely on long-term memory to support complex task execution, user personalization, and domain adaptation. Meanwhile, emerging access-control mechanisms for LLM agents are being explored to block policy-violating requests and prevent misuse. We reveal a novel attack surface arising from agent memory operations: prohibited content that would trigger access control can be fragmented across interactions, stored in long-term memory in benign-appearing form, and later reconstructed through memory retrieval without appearing explicitly in the final user query. We propose FragFuse, the first attack that enables unprivileged users to bypass agent access control by exploiting this temporal channel introduced by long-term memory. FragFuse operates in three stages: (1) identifying rejection-responsive fragments via black-box adaptive querying with fragment masking; (2) injecting these fragments into memory using marker carrier queries; and (3) retrieving and fusing the stored fragments through a follow-up attack query. Although FragFuse can be instantiated manually for individual agents, we further develop a surrogate-based optimization scheme that tunes fusion instructions and marker designs, enabling automated attack generation without violating the attacker's threat-model assumptions. We evaluate FragFuse across four representative agent settings and task domains, covering three state-of-the-art agent access-control mechanisms. FragFuse achieves an average bypass success rate of 86.3% and an average end-to-end harmful task success rate of 41.1% across all settings, with only 4.4% average task-success degradation compared with configurations without access control. We also show that alternative defenses, including state-of-the-art prompt-injection detectors and perplexity detectors, do not effectively address this attack.

The adoption of AI agents is increasing rapidly. Terminal AI agents, i.e., AI agents that run in terminal environments, are a widely used type of AI agents. Terminal AI agents rely heavily on shell command execution to interact with the host systems. They adopt a three-list command-gating mechanism to mitigate security risks introduced by command execution, with denylists serving as the load-bearing component. However, modern operating systems often ship a large, ever-expanding set of shell commands with complex functionalities. Our observation is that even a built-in denylist of Claude Code, well-maintained by its developers, can overlook bypass commands that invalidate its effectiveness. Such negligence leads to fragile command denylists that cannot even block operations that practitioners expect them to block. This paper presents the first systematic characterization of command denylist fragility in terminal AI agents. The paper formalizes the command denylist fragility problem and proposes an LLM-driven pipeline, ShellSieve, to detect such fragility. It prompts the LLM to propose possible bypasses and iteratively repairs them using feedback from a validator that executes them in a sandbox. In the evaluation, we applied ShellSieve to 1,709 real-world command denylists (containing 13,332 denylist rules) collected from GitHub. The evaluation shows several key findings, including that 69.0–98.6% of the denylists are fragile, that this fragility occurs consistently across projects and agents, and the validity of several possible root causes for this fragility. Our pipeline and findings will hopefully facilitate future research and practice regarding the command denylists used by AI agents.

Answering statistical queries over sensitive data under differential privacy (DP) is a common task in many settings, including databases, mobile computing, and data markets. In these scenarios, multiple analysts may issue the same query, while receiving answers generated under different privacy budgets due to differences in trust levels or willingness to pay. Existing approaches for such multi-tier DP queries either incur excessive cumulative privacy loss or suffer from suboptimal utility. In this paper, we propose a framework for multi-tier DP query release that simultaneously bound the cumulative privacy loss by the maximum privacy budget among all queries and achieve optimal utility comparable to that of single-tier mechanisms. Our framework applies to different classes of DP mechanisms. For noise-adding mechanisms (e.g., count queries with the two-sided Geometric mechanism in the curator model), we develop a general solution based on the characteristic functions of noise distributions. For other mechanisms (e.g., count queries under the local DP model with the Subset mechanism), we design mechanism-specific primitives for budget transformation and introduce a template-based strategy that attains optimal utility across different privacy regimes. Experimental results demonstrate the effectiveness of our framework.

Adversarial attacks on large language models have limited practical impact despite extensive research. Optimization-based attacks such as Greedy Coordinate Gradient (GCG) (Zou et al., 2023) produce high-perplexity, incoherent suffixes that existing defenses easily detect (Bengio et al., 2024). Moreover, attempting to enforce coherence constraints during optimization often prevents the attack from successfully eliciting the specific targeted response, resulting in low success rates against robust models. Conversely, attacks that maintain coherence often alter the semantic intent of queries; when the model complies with these altered queries, responses fail to address the adversary's original goal. In this work, we introduce Greedy Coordinate Diffusion (GCD), a novel framework that efficiently generates adversarial attacks against safety-aligned models while maintaining low perplexity and high semantic adherence to the adversary's original intent. GCD leverages the generative priors of discrete diffusion language models to guide the search for adversarial suffixes that achieve semantic coherence and adherence. Unlike GCG, GCD does not require direct gradient access, allowing it to operate in a gray-box setting. We show GCD achieves highest ASR while remaining competitive on response-quality scores, and that the constructed adversarial prompts are detected at lower rates than other methods by perplexity-based and guard-model filters.

Mobile Ad-hoc Network (MANET) is temporary and dynamic network topology, wherein nodes are mobile in nature and distributed randomly in a network area. In MANET, nodes cooperate with each other to operate and forward data through multihop communication between source and destination. MANET is exposed to different types of attacks due to absence of central administration. However, some nodes decline to cooperate, misbehaves and appears to be malicious affecting network functionality and connectivity. Providing security and identifying malicious node has become one of the challenging research topics in MANET. Black-hole attack is considered to be most popular attack that degrades the overall network performance. Black-hole node falsely advertises the shortest path to destination intentionally to disrupt the network communication resulting in packet drop. In collaborative black-hole attacks, multiple black-hole nodes cooperate and launch attacks in order to degrade network reliability. In this article we propose a Lightweight technique to detect and isolate Collaborative Black-Hole attacks (LW-CBH) by enhancing existing AODV routing protocol. In this scheme a timer based baiting process and reverse tracing setup is used to detect malicious node through control status message in MAC layer which are Reply Sequence (R-SEQ) and Code Sequence (C-SEQ) message of connected dominated set of nodes. However existing AODV routing protocol fails to detect malicious node during dynamic topology changing in MANET. Simulation of proposed technique is performed using discrete event simulator tool NS-2.35. The simulation results are evaluated for throughput, packet delivery ratio, average end-to-end delay and normalized routing overhead.

Agentic systems are increasingly integrated with geographic information systems (GIS), where multi-agent coordination enables complex conversational and spatial analysis but introduces security risks. This work presents a security-oriented framework for risk identification, evaluation, and mitigation in a multi-agent GIS system while maintaining adaptability to broader agentic architectures. We test the agentic system of a commercial geospatial partner while developing a modular state-machine-based orchestration framework that abstracts agent behavior into reusable components. We evaluate robustness using a red-teaming framework with an adaptive attacker LLM and a deterministic judge that produces binary outcomes with supporting rationales across multi-turn attacks. We further improve resilience with a prompt optimization framework that treats prompts as structured signatures and injects adversarial demonstrations, enabling systematic security improvements without degrading task performance.

Model stealing attacks, where adversaries create high-fidelity surrogate models, are a significant threat to the intellectual property of machine learning services. Conventional wisdom suggests these surrogates could provide adversaries with economic leverage comparable to the original service providers. This paper challenges this assumption by evaluating model stealing attacks beyond mere fidelity to the target model. Because query-based extraction provides only partial supervision of the target's input-output behavior, the surrogate is not uniquely identified: many near-optimal surrogates can achieve comparable fidelity while differing in deployment-relevant properties. Instead of performing a classic learning-based model stealing attack, we compute the Rashomon Set (i.e., the set of almost-equally-accurate models) of surrogate models, and evaluate its diversity using multiplicity metrics (ambiguity, discrepancy, and Rashomon Capacity) and group fairness metrics. Across tabular, medical imaging, and NLP tasks, our experiments on real-world datasets reveal that despite exhibiting similar fidelity to the target model, surrogate models can display significant variances in other critical performance metrics. These findings cast doubt on the presumed equivalence between high-fidelity surrogates and the target model in practical deployment scenarios.

We experimentally demonstrate that it is possible to generate unique, repeatable, and device-specific signatures suitable for use as Physical Unclonable Function (PUF) responses in commercial off-the-shelf (COTS) DRAM chips by leveraging simultaneous multiple-row activation (SiMRA). Based on a rigorous experimental characterization of 112 modern DDR4 DRAM chips (from 10 modules), we introduce SiMRA-PUF, the first DRAM-based PUF that uses SiMRA-generated signatures as PUF responses. We analyze SiMRA-PUF in terms of reliability, uniqueness, and evaluation latency for varying numbers of simultaneously activated DRAM rows (i.e., 2, 4, 8, 16, and 32), DRAM chip density & die revision, and evaluate how temperature affects the similarity of SiMRA-generated responses. Among our 8 key experimental observations, we highlight two major results. First, SiMRA-PUF provides average intra-Jaccard indices of 89.02%, 89.81%, 93.03%, 94.06%, and 94.86%, and average inter-Jaccard indices of 3.98%, 2.37%, 3.44%, 2.92%, and 3.24% for 2-, 4-, 8-, 16-, and 32-row activations, respectively, showing that SiMRA-generated signatures are both repeatable within a device and unique across devices. Second, 2-row activation-based SiMRA-PUF provides 5.75% lower evaluation latency than the state-of-the-art DRAM-based PUF. We open-source our infrastructure and datasets at https://github.com/CMU-SAFARI/SiMRA-PUF.

This paper presents an empirical analysis of the Web3 security landscape over the four-year and three-month period from 1 January 2022 to 27 March 2026. The dataset combines 23,818 public audit findings produced by 22 independent security firms with 218 real-world exploit incidents documented by rekt.news, representing aggregate losses of approximately US$7.76 billion. We report three central findings. First, the distribution of audit findings (by severity, category, and technology stack) is substantially stable across the observation window, with the Critical-plus-High share remaining within a 15-17% band in every complete year. Second, the categorical distribution of realised exploit losses does not correspond to the categorical distribution of audit findings: private-key compromise, phishing, and social-engineering vectors account for approximately 49.6% of cumulative losses yet represent a negligible share of published audit findings. Third, realised losses exhibit extreme concentration: the eight largest incidents account for 50.6% of cumulative dollar losses and the twenty largest for 71.4%, a distributional shape inconsistent with Gaussian assumptions. Throughout, we adopt the analytical convention that audit outputs and exploit outputs describe different populations and present the two datasets in parallel rather than as directly comparable samples.

Indirect prompt injection attacks hijack LLM-based agents by embedding malicious instructions in third-party data that the agent retrieves during task execution. Existing defenses report near-zero attack success rate on static benchmarks, yet recent adaptive evaluations show that these results collapse once the attacker is allowed to optimize against the deployed defense. In this work, we trace this collapse to two failure modes. First, existing defense methods are confined to recognizing specific attack patterns, rather than assessing whether the intent of every embedded instruction is relevant to the user task. Second, training-based defenses, which otherwise offer the strongest safety-utility trade-off, assemble their adversarial examples from a handful of hand-crafted templates, and the resulting defender fails to generalize outside that narrow strategy distribution. To address these gaps, we propose RETA, a training-based method that grounds defense decisions on the user tasks rather than attacker-controlled data. At each tool-output step, the defender undertakes chain-of-thought reasoning verifying that its actions are consistent with the user task. Leveraging red-teaming, a simulated attacker synthesizes adversarial training data and receives a dictionary-learning diversity reward, achieving broad coverage of injection-reformulation strategies. Together, these allow the defender to be optimized via multi-objective reinforcement learning and achieve better safety-utility trade-off. Across six black-box adaptive attacks, RETA keeps every per-attack ASR below 10%, with average ASR of 2.92% and 3.75% on the two target models, while preserving most utility under attack and on clean inputs.

Skills are becoming the capability layer through which LLM agents turn plans into actions, but their use introduces security risks such as data leakage, unauthorized operations, and tool misuse. Existing vetting usually evaluates each skill in isolation, while real agent tasks often invoke multiple skills in a shared execution context. This creates Skill Composition Risk (SCR): a skill that appears benign alone can become harmful when its outputs, trust signals, authorization cues, or side effects influence later invocations along an activated path. We introduce SCR-Bench to evaluate this risk in controlled, sandboxed skill environments. Rather than relying only on textual intent or surface behavior, SCR-Bench records downstream state changes and path-level outcomes across composed skill executions. It contains three sub-benchmarks: SCR-CapFlow for capability-flow composition, SCR-TrustLift for trust-transfer composition, and SCR-AuthBlur for authorization-confusion composition. Across SCR-Bench, composed paths expose risks that are largely absent under isolated evaluation. In SCR-CapFlow, attack success rate reaches 33.6 percent under composition, compared with near-zero isolated baselines. In SCR-TrustLift, attack success rate exceeds 96.5 percent on four of five backends. In SCR-AuthBlur, the risky-approval rate increases by 71.8 percent relative to the L0 isolated baseline under the L1 context setting. These results show that agent skill security should be assessed at the level of activated paths rather than isolated artifacts. SCR and SCR-Bench provide a foundation for path-aware risk evaluation and defense in LLM agent skill ecosystems. Benchmark: https://github.com/saint-viperx/SCR_Bench.

Network scanning is a critical preliminary step for most adversaries to gain essential information before launching cyber attacks. Moving Target Defense (MTD) based on IP shuffling has emerged as a proactive defense strategy to counteract these reconnaissance efforts. Unlike static, reactive defense techniques, IP shuffling introduces randomness by dynamically reassigning network addresses, making it more challenging for attackers to identify and track targets. However, current IP shuffling methods face three key challenges: 1) limited scalability across different network topologies, 2) inherent reconfiguration overhead even in the absence of an active attack, and 3) the need for large-scale unused address blocks. To address these issues, we propose LSTM Look-ahead Moving Target Defense (LLM). Our approach is the first attempt using a Long Short-Term Memory (LSTM) network to predict future target addresses that attackers will likely scan. Ensemble learning is used to improve robustness to different scanning behaviors. We introduce a dynamic mutation mechanism to enhance adaptability. Compared to the baseline mutation strategy, LLM performs better in both security and overhead.

Air fingerprinting infers application activity by sniffing metadata from cellular control channels. 5G encrypts these channels, breaking the attack chain that prior attacks depend on. This paper reveals a physical-layer side channel that bypasses encryption: under the link adaptation mandated by the cellular communication standard, the uplink Modulation and Coding Scheme (MCS) remains stable, so the number of Physical Resource Blocks (PRBs) occupied by a transmission accurately reflects the IP packet length. Combined with the uplink control channel that carries downlink information, an attacker can reconstruct a bidirectional traffic profile. This bidirectional information recovery can be achieved simply by observing the uplink spectrum, without decoding any channel. Building on this side channel, we design Crosshair, a passive three-step attack. First, a blind extraction stage recovers the uplink physical channel occupancy from raw IQ samples via energy detection, reconstructing bidirectional traffic from uplink spectrum. Second, we design a data augmentation method that synthesizes spectral profiles across diverse channel conditions, eliminating the need for prior knowledge of the communication environment. Third, cross-modal alignment embeds the spectral and IP domains into a shared space, enabling new applications to be enrolled from a collected IP trace alone. Extensive experiments on a 5G NR testbed demonstrate the robustness and precision of Crosshair: it outperforms the State-of-the-Art (SOTA) physical layer fingerprinting method in application recognition accuracy, and maintains high accuracy in cross-MCS scenarios.

Targeted advertising systems can pair audiences selected by advertisers with ad units that expose visible user actions. When an interaction remains linked to the campaign that elicited it, the advertiser may receive an observation tied to a user rather than only an aggregate report. We model that channel as a noisy oracle for attribute inference. The model separates targeting predicates, exposure, interaction, and disclosure. These boundaries capture the gap between eligibility and delivery, and the gap between interaction and advertiser visibility. We build a reproducible benchmark using synthetic populations calibrated with public data, each with known sensitive labels. A generated campaign semantics layer provides topic variants and response priors. The simulator generates the ground truth, event traces, disclosed observations, and metrics. The evaluation compares Bayesian, supervised, positive and unlabeled, and adaptive attacks under common campaign and disclosure definitions. The final evaluation uses four topic variants, seven simulator seeds, and two interaction settings. Repeated campaigns with identity exposure produce measurable but bounded inference signal. At $160$ campaigns, Bayesian and supervised attacks reach about $0.64$ AUC in the main setting and about $0.65$ AUC in the higher interaction setting. Disclosure policy is the strongest control. Aggregate reporting removes the evaluated oracle input tied to users. Type filtering and randomized disclosure reduce the released signal. The result is a model, artifact, and defense evaluation method for privacy in interactive targeted advertising. The code is available at https://github.com/P-HOW/Interactive-Ad-Oracle.

Vision-Language-Action (VLA) models enable end-to-end robot control and have garnered widespread attention. However, the memorization of training data inherent to VLA, coupled with the high cost of robotic data acquisition, raises serious concerns regarding data privacy leakage and intellectual property infringement. Membership inference attacks (MIAs) aim to determine whether a given sample belongs to the training set. While representing a significant privacy threat, this attack remains underexplored in the context of VLA models. To bridge this gap, we propose VLALeaks, which is based on attention discrepancies in VLA models. We reveal, for the first time, the privacy vulnerabilities of VLA models. Specifically, it comprises a two-stage process: (1) membership feature extraction, and (2) attack model construction. Experimental results across multiple VLA benchmarks demonstrate that VLALeaks readily reveals membership information and achieves optimal attack AUC and TPR@1%FPR, highlighting the privacy vulnerabilities in current VLA model deployments. Our work is the first systematic study of MIAs on VLA models, aiming to provide insights for secure and trustworthy VLA models.

We study the task of CVE-conditioned exploit generation, where a model drafts proof-of-concept (PoC) exploits given software vulnerability context. We adopt a data-centric approach, constructing a high-quality dataset via multi-stage preprocessing and introducing a scalable evaluation framework with LLM-as-judge and fine-grained rubrics. Under this unified setup, we benchmark 17 large language models across 8 evaluation criteria, providing systematic insights into their zero-shot capabilities. We further show that a compact 8B open-weight model, when fine-tuned on curated data, achieves over 42.5% improvement in exploit quality and rivals some proprietary models when combined with simple test-time rejection strategies. Our results highlight the importance of data quality, structured supervision, and evaluation design for reliable exploit generation, suggesting that these factors can be as critical as model scale in adapting LLMs to cybersecurity tasks.

Previous FPSI works have demonstrated a linear scaling with the distance threshold $\delta$, while some recent works have achieved a poly-logarithmic dependence on $\delta$. However, these protocols either support only the $L_\infty$ distance, or they support general $L_{p\in[1,\infty]}$ distances but rely on expensive additive homomorphic encryption (AHE). Achieving exact logarithmic dependence on $\delta$ for general $L_{p\in[1,\infty]}$ distances without relying on costly AHE would constitute a theoretical breakthrough in optimal threshold scaling and a practical advance toward scalable FPSI applications. In this work, we present new FPSI protocols for $L_{p\in[1,\infty]}$ distances that are entirely built from oblivious transfer (OT) and symmetric-key primitives. We propose FPSI protocols based on both the apart and the separate assumptions, which are applicable to low- and high-dimensional settings, respectively. Our constructions achieve strictly logarithmic complexity in $\delta$, which is optimal in the sense that distinguishing all values in an interval of length $O(\delta)$ necessarily requires $\Omega(\log \delta)$ bits of information. Our core idea is to perform fuzzy matching via prefix representation and interactively determine the correct prefix using equality conditions. To this end, we propose a suite of new components that can be implemented efficiently using only OT and symmetric-key operations. We implement our FPSI protocols and compare them with the state-of-the-art FPSI protocols for $L_{p\in[1,\infty]}$ distance. Experiments show that our protocols outperform the prior state-of-the-art by up to $43.7\times$ in runtime and $31.3\times$ in communication.

Indirect prompt injection (IPI) is a major security threat to LLM-powered agents. Thus, a growing body of work has proposed a variety of defensive approaches against IPI which can be grouped into three broad categories: prompt-level, filter-based, and system-level designs. However, commonly used benchmarks for evaluating defense, such as AgentDojo, are inherently static, generating a fixed distribution of IPI attacks. Consequently, a defense can score well on them without being robust to adaptive threats. We introduce AutoDojo, a generative benchmark built on AgentDojo and AgentDyn that generates IPI adaptively for a given agent and defense. It supports six task suites across the two benchmarks, covering banking, communication, travel, shopping, coding, and everyday-assistant domains. AutoDojo optimizes injections under a strict black-box setting, observing only whether a candidate injection succeeds, and can readily integrate, and often improve on, any existing black-box attack. Across ten defenses and five target models, AutoDojo demonstrates that standard static benchmarks often significantly overestimate defense efficacy. Moreover, we show that most existing defenses either sacrifice considerable utility or are insecure. Finally, we demonstrate that attack success interacts with task specification, with under-specified tasks particularly vulnerable. AutoDojo is available at https://github.com/xhOwenMa/AutoDojo

Moving Target Defense (MTD) is a proactive security strategy designed to enhance cyber-resilience by dynamically altering system parameters, thereby preventing adversaries from acquiring the critical information needed to execute stealth attacks. In this paper, we consider the case in which the operator modifies the admittance of branches to enable MTD, and focus on the problem of effectively protecting the system with fewer number of branch admittance modifications and shorter computational time. Specifically, we identify the ineffectual branches whose admittance modification do not contribute to the improvement of MTD effectiveness via theoretical analysis. Building on these insights, we propose the Bus Traversal-based MTD (BT-MTD), which is a bus-oriented algorithm that traverses over the buses of the network according to analytically derived guidelines. The performance of the BT-MTD is evaluated and compared with four existing strategies on standard IEEE test systems, demonstrating its robustness and superior performance in effectiveness, efficiency, and computational cost. The code of BT-MTD is available at: https://github.com/YJY101/BT-MTD.

Production LLM applications commonly stack a regex filter in front of model-side alignment; prior work found no measurable coverage gain from adding a live Gemini backend behind an active regex filter. We ask whether that ceiling holds when the corpus is designed to bypass the regex. We introduce $L_5$-no-regex – identical to $L_4$-real (Gemini-2.5-flash, token-budget cap, rate limit, output scrub) but with the nine-pattern filter disabled – and evaluate it against $N{=}45$ adversarial probes across three sub-corpora (carry-forward, regex-bypass, alignment-isolate), amplified by Gemini paraphrase and PAIR to ${\sim}1{,}555$ probe-run pairs over $N{=}5$ replications. Under the primary substring classifier, H1 is refuted: $L_5$ block rate is $0,%$ across all five OWASP LLM Top-10 categories ($\Delta\text{pp}{=}0$ vs.\ $L_0$, $p{=}1.00$; Wilson upper bound ${<}5,%$). A secondary LLM-judge metric on PAIR variants shows $56$–$100,%$ block rates ($p{<}0.01$), revealing alignment does respond to adversarially-framed probes – but produces refusals too nuanced for substring matching. The sub-corpus differential prediction is not supported ($p{=}1.00$). Alignment's contribution is metric-dependent: on natural-language harmful-request probes, it adds zero observed coverage beyond the regex; on adversarially-framed variants, an LLM judge detects refusals the substring classifier misses. The locked corpus, mutation artifacts, and export scripts are released for replication.

LLM supervision systems, namely input/output moderation filters and jailbreak detectors, are the primary safeguard against misuse in deployed AI applications, yet existing benchmarks are often vendor-biased, omit cost and latency, and rarely compare specialized guardrails against repurposed generalist LLMs. We present BELLS-O (Benchmark for the Evaluation of LLM Supervision Systems, Operational), the first independent operational benchmark of LLM supervision systems. BELLS-O evaluates 28 systems from 17 providers: every major specialized guardrail (e.g., LlamaGuard-4, ShieldGemma-2, Lakera Guard) and frontier generalists repurposed as supervisors (e.g., GPT-5.4, Claude Sonnet 4.6, Grok-4.1), jointly on detection rate, false-positive rate, latency, and monetary cost. We cover input/output moderation across 11 harm categories and jailbreak detection across 13 attack techniques, using in-house datasets built from handcrafted prompts, expert-curated samples, and quality-controlled synthetic generation. To suppress latent generator fingerprints in synthetic data, every generated sample is paraphrased. Mapping the Pareto frontier reveals use-case-dependent tradeoffs. On content moderation, specialized supervisors are operationally dominant: top systems match frontier LLMs on detection (~95% vs. 94%) at comparably low false-positive rates (<=2%), while running 5-10x faster and ~10x cheaper. On jailbreak detection, the tradeoff shifts: frontier LLMs achieve higher detection and lower false-positive rates but at 10-50x higher cost and 5-10x higher latency. We release the benchmark, framework, leaderboard, and datasets as the first vendor-neutral basis for selecting safeguards under real deployment constraints.

Building automation and control systems integrate heating, ventilation, air conditioning, lighting, sensing, and management functions through specialized communication protocols. While this integration enables flexible building operation, it also creates complex cyber-physical environments that are difficult to inspect, secure, and explain to new analysts. This paper presents a practical security and human-centered case study of a BACnet/IP building automation testbed with DALI lighting infrastructure, investigated during a domotics-oriented cybersecurity hackathon in Thun, Switzerland in April 2026. The study combines network-oriented enumeration, object-level inspection, physical rack analysis, and reflective HCI analysis of tool-supported learning. Using Yabe and BACteria, the work documents observable BACnet services, reconstructs structured object hierarchies, identifies room-level lighting-control paths, and maps BACnet objects to DALI group-level infrastructure. The analysis emphasizes that BACS assessment is not only a technical protocol task: it also requires usable tool interfaces, physical observability, interpretable naming conventions, and safe mental models for command priorities. The paper contributes a compact case study of BACnet/DALI exploration in an educational testbed and discusses implications for cybersecurity education, human-centered security tooling, and responsible experimentation in cyber-physical building environments.

Document-to-LLM applications typically read uploaded PDFs by first translating them into text through an extraction layer whose output is not ordinarily exposed to users. We show that this layer enables split-view PDFs: one document can have two semantic views before model reasoning. By mining specification-permitted or implementation-tolerated representation gaps at the PDF render/extract boundary, we instantiate 25 extraction gaps (EG) in which extractors return attacker-controlled or extractor-dependent text while the rendered page shows benign or different content. The gaps form four families: semantic overrides, hidden semantic injection, reading-order splits, and font-decoding splits, and 14 gaps have no exact path/mechanism-level match in prior PDF-to-LLM attacks. We evaluate these gaps on 16 PDF processing stacks and 7 commercial LLM services. Each gap causes render-extract divergence on at least one stack. Under a gap-level exposure criterion, every evaluated service exposes at least one gap, with 12/25 to 21/25 exposed gaps. Exposure is driven mainly by the ingestion stack – not model identity alone. We further show that tested safety filters cover only selected hidden-text constructions. To support triage, we develop a static screening scanner whose rules trigger on all 25 benchmark gaps, and discuss dual-view consistency as a longer-term defense direction.

Agentic large language model (LLM) systems can now execute actions, not only produce text. When model outputs trigger privileged operations such as shell commands, browser automation, or external tool calls, the security problem shifts from alignment alone to system configuration and structural design. We analyze OpenClaw, a self-hosted multi-agent system in which LLM outputs can execute commands and interact with tools and services. We measure compromise probability, boundary failures, privilege drift, and how these metrics change as attacker capability increases. With one agent, the compromise probability is 0.24. With seven agents, when the system executes an action, the compromise rises to 0.86 if any single agent proposes it. The models do not change; the increase comes from output aggregation. Prompt injection propagates instability across the system. Attack surface entropy increases from 0.42 to 0.71, indicating a broader distribution of exploit paths. The mean privilege drift increases from 0.03 to 0.21, indicating unintended authority gain. Positive escalation curvature of 0.08 indicates that privilege grows faster as attacker capability increases. Defensive controls, including policy gating and execution filtering, reduce compromise probability by 0.10, boundary failures by 0.10, and privilege drift by 0.02, all statistically significant at p < 0.0001. The system remains sensitive, but the mitigation impact is measurable. Injection mitigation success differs across models: 0.37 for GPT-5.2, 0.35 for Llama-4-Maverick, and 0.31 for DeepSeek-R1. When execution can be triggered by any single agent, the most vulnerable agent determines system exposure. Mitigations slightly reduce task utility from 0.93 to 0.89 and increase median latency from 420 ms to 468 ms.

Internet of Things (IoT) and Cyber-physical systems (CPS) increasingly rely on continual learning (CL) to adapt to evolving environments, device heterogeneity, and concept drift, thereby improving overall utility. While continual adaptation is essential for long-lived IoT deployments where data patterns evolve, it also introduces new security vulnerabilities. In particular, backdoor attacks can exploit incremental updates, replay buffers, and representation reuse to implant persistent malicious behaviors that remain dormant during normal operation but activate upon specific triggers. In this paper, we present a backdoor attack in continual learning used in IoT/CPS systems. To this end, we formalize an IoT/CPS-specific threat model, analyze why continual learning amplifies backdoor persistence in IoT pipelines, and evaluate our technique under varying conditions. Our analysis highlights critical open challenges in securing lifelong learning in IoT/CPS and industrial IoT (IIoT) environments, as well as the need for heightened security controls.

Submodular maximization subject to a knapsack constraint (SMK) is a fundamental problem in discrete optimization, with wide-ranging applications in machine learning and related fields. As these applications increasingly involve sensitive individual data, there is a growing need for high-utility algorithms that provide formal privacy guarantees. In this work, we study the SMK problem under differential privacy, considering both monotone and non-monotone objective functions. For monotone objectives, we propose a differentially private algorithm that achieves the optimal $(1-1/e)$-approximation ratio while significantly improving both additive error and query complexity over prior work. We also present a more efficient algorithm for the same setting, achieving a $1/2$-approximation. For non-monotone objectives, we introduce, to our knowledge, the first differentially private algorithm with provable guarantees, achieving a $1/4$-approximation in expectation and an additive error comparable to the best known for monotone objective functions.

Auctions are now central to blockchain markets, settling NFT sales, token launches, DeFi liquidations, and arbitrage opportunities. Each on-chain bid is a public transaction whose inclusion is decided by a single consensus proposer per block. The proposer can observe pending bids, exclude competitors, and submit bids of their own, breaking the fairness guarantees of classical sealed-bid auctions. To enable latency-sensitive sealed-bid auctions in blockchain settings, we formalize four properties – each necessary to prevent a concrete attack – and design a protocol achieving all four: hiding bid contents, existence, and bidder identity until reveal (Hiding); counting all timely honest bids and rejecting late adversarial bids (Simultaneous Release); preventing silent withdrawal of committed bids (No Free Bid Withdrawal); and charging on-chain fees only to winners (Auction Participation Efficiency). Our protocol uses a timestamping oracle (instantiated with a committee of 2f_ts+1 timestampers) and a censorship-resistant inclusion predicate (instantiated using a FOCIL-based inclusion list), with only the winning bid settled on-chain. Our construction relies on two zero-knowledge proofs: an eligibility proof that anonymously proves deposit membership to the timestamping committee, and an auction proof that binds a bid to a specific auction for the inclusion list committee. We implement both using Groth16 over BN254 with Poseidon hashing in arkworks/Rust: the auction proof generates in 13 ms and verifies in under 1 ms; eligibility proofs for Merkle trees up to 2^32 bidders generate in 47-159 ms and verify in about 1 ms. Together, this yields a sealed-bid auction primitive practical for high-value, time-sensitive blockchain settings.