Papers

Post-quantum migration relies on hybrid X.509 certificates, which carry post-quantum material alongside the classical so existing verifiers still work. Several designs place it where a verifier may ignore it, so the classical path decides. We tested eight open-source path-validation stacks over seven independent codebases, one in two builds: nine configurations over six certificate profiles. On their default paths, every stack that parsed a separable hybrid certificate accepted it. Invalidating the post-quantum evidence in each separable scheme, leaving the classical evidence valid, changed no verdict in any of the 27 cells: none distinguished sound post-quantum evidence from destroyed. Four stacks verify post-quantum signatures elsewhere on the same path, so immature support does not explain it. Two stacks implement the checks the schemes specify, neither on its default path, and no document defines the interface between them: one carrying a relying party's hybrid requirement, an operational policy, into path validation and reporting which kind of acceptance resulted. We contribute a specification-derived model, this test, and a policy-parametric contract pairing a policy input with a labelled result. Revoking a bound post-quantum certificate changes no verdict in any of the nine configurations, because none consults it; the labelled result makes it visible to operations.

Large language models fine-tuned for network intrusion detection emit single-point predictions without statistical validity guarantees. Conformal prediction supplies a finite-sample coverage guarantee, but a threshold calibrated on clean traffic fails once an adversary perturbs controllable network features. We demonstrate this failure across three intrusion detection benchmarks and propose traffic-aware conformal prediction, which calibrates on traffic drawn from the perturbation mechanism an attacker is expected to use and provably restores coverage whenever that mechanism is known and can be sampled. A stronger, adaptive attacker that queries the target model's own score can still degrade this matched-calibration guarantee. We address this second threat model by excluding attacker-controllable features and their deterministic descendants from the scored representation, and prove that this yields an exact, pathwise coverage guarantee rather than a probabilistic bound. Across three independently fine-tuned language model architectures, this representation remains completely unchanged under every evaluated attack attempt, at a quantified seven-to-fourteen-point cost in clean accuracy relative to the unrestricted feature set.

AI coding agents powered by LLMs are increasingly integrated into real-world software development, where they generate, edit, and execute code with autonomous access to local files and tools. Coding agents inherit security risks from both the LLM backbone, where adversarial prompts, poisoned training data, and backdoor triggers can cause models to emit insecure or attacker-chosen code, and their agentic architecture, where tool-using autonomy enables induced misuse of external APIs, data exfiltration, and persistent compromise of development environments. This paper presents a systematic evaluation of malicious issue requests against state-of-the-art coding agents (Cursor, Claude Code, and Codex Desktop), powered by two major model families (OpenAI GPT-5.3 Codex/GPT-5.4 and Anthropic Sonnet 4.6). Our novel benchmark IssueTrojanBench contains malicious issues that are constructed based on four novel attack categories (i.e., embedded as malicious instructions in issues), six delivery vectors (e.g., PDF, or issue comment), and further augmented by perturbations. Our results reveal critical vulnerabilities in the as-deployed modern coding agents, i.e., 66.5% of the malicious issues from IssueTrojanBench penetrate all the guardrails (agent- and LLM-level) of coding agents. Our further analysis shows that rejection is almost entirely from LLMs rather than the agent frameworks, with GPT models broadly vulnerable and Sonnet 4.6 exhibiting more selective, risk-aware blocking of high-impact actions. Our evaluation also highlights that the current agent-level defense strategy offers limited additional protection for coding agents. Our findings highlight the urgent need for stronger agent- and model-level safety mechanisms to protect AI coding agents.

Large language models increasingly use search tools to retrieve up-to-date information, introducing a new attack surface in which retrieved documents can be manipulated. This risk is amplified by the development of generative engine optimization, which can make selected content more likely to be retrieved, cited, and adopted by models. Existing fact-verification benchmarks and evaluation frameworks do not provide the controlled evidence environments needed to assess robustness against GEO poisoning. We therefore propose GPE, which consists of a multi-domain fact-verification benchmark and an evaluation framework for controlling evidence sources and poisoning ratios. Experiments across multiple verification methods and poisoning attacks demonstrate that GPE exposes robustness degradation and efficiency trade-offs that cannot be observed through clean evaluation alone, confirming the need to evaluate fact verification under adversarial evidence environments.

Differential privacy under Continual Observation (CO) quantifies the loss in privacy that occurs when outputs generated using a stream of sensitive input data are published in the online setting. In this paper, we consider a more stringent notion of privacy compared to prior work wherein an individual's participation may shift the entire stream by a time-step. We define a new notion of edit-neighboring streams that captures this scenario. Our findings are as follows. First, we prove that on a stream of length $T$, every additive-noise mechanism incurs error $\tilde{\Omega}(\min\{T^{1/3}/\varepsilon^{2/3}, T\})$ when required to be $\varepsilon$-DP under CO for edit-neighboring streams. This includes state-of-the-art continual counters constructed via the factorization mechanism that in the standard neighboring setting incur only polylogarithmic additive error. Second, we construct the first mechanisms with polylogarithmic additive error for our more stringent notion of privacy. We show that we can recover the same additive error as in the standard notion of privacy albeit with worse constant coefficients for both arbitrary input streams and sparse streams. Third, we show that the notion of edit-neighboring streams inhabits a `sweet-spot' in terms of generality and additive error incurred. More precisely, we show that the even more general notion of prefix-sum neighboring streams—which arises naturally in reductions for problems under CO—must incur additive error scaling as $\tilde{\Omega}(\min\{T^{1/3}/\varepsilon^{2/3}, T\})$ for any mechanism that is $\varepsilon$-DP under continual observation. Finally, we show empirically on synthetic data that when compared with prior work, our mechanism achieves a superior trade-off between the success probability of a simple distinguishing attack, and the additive error incurred by the respective mechanisms.

This work presents LeakyLMs, a set of attacks that leak proprietary model, architecture, and deployment information from production language models. LeakyLMs is the first to demonstrate that key model and deployment details can be inferred using only token generation timing, even when interacting through remote APIs. LeakyLMs introduces two core attacks. The first attack targets inference optimizations and deployment strategies. For example, our attack detects whether a provider uses speculative decoding, a widely deployed inference-time optimization, and further identifies the context length of the draft model used in the pipeline. Our measurements show that Google Gemini Flash 2.5 uses speculative decoding with a draft context window of approximately 128K tokens. The second attack recovers key architectural properties, including the number of transformer layers, hidden dimension size, and number of attention heads. To achieve this, LeakyLMs builds a detailed and accurate model of token-generation timing on modern NVIDIA GPUs, characterizing how latency scales with model configuration and hardware parameters. The attack then performs a search over the architecture space using this timing model. In experiments with Llama models, the near-correct architectural configuration appears in the top-10 guesses more than 90% of the time.

LLM-based coding tools enable non-expert users to generate routine automation scripts that may enter enterprise workflows without meaningful security review. This study examines that risk directly. Code was collected from ChatGPT, Microsoft Copilot, and Google Gemini using identical prompts across three automation domains. Claude Code performed a standardized vulnerability review. Each identified vulnerability was scored using CVSS v3.1 and mapped to the OWASP Top 10:2021 and the MITRE ATT&CK frameworks. Every script contained exploitable vulnerabilities. Nine of the 17 identified vulnerability classes appeared in code from all three models, while 14 of the 17 vulnerability classes appeared in at least two models. The weighted CVSS scores across platforms differed by less than 10%. The risk is not tied to any particular model but rather to the task category. Organizations should therefore ask not which tool to trust, but instead whether LLM-generated automation code should be deployed without review.

Security protocols verification relies on formal tools such as ProVerif and OFMC. This study evaluates whether large language models (LLMs) can perform comparable analysis. We test GPT and DeepSeek in chat and reasoning modes over three runs on 130 obfuscated AnB/AnBx protocols covering 388 security goals, scored against ProVerif and OFMC. Each provider uses a single model in both modes, switching reasoning on and off, so both contrasts isolate reasoning itself. Chat models achieve 72.7% recall at 27.3% precision for GPT and 69.3% recall at 27.2% precision for DeepSeek. Reasoning models reverse this trade-off, reaching 66.5% precision and 54.5% recall for GPT and 45.4% precision and 57.3% recall for DeepSeek. Enabling reasoning lifts precision from 27.3% to 64.8% for GPT and from 27.2% to 44.4% for DeepSeek on the consolidated verdict. The goal set is imbalanced, with 89 vulnerable goals against 299 secure ones; a trivial always-secure predictor scores 77.1% accuracy, which only GPT reasoning exceeds. All models perform worst on authentication goals: reasoning models detect well under half of injective and non-injective agreement attacks, whereas chat models over-flag them at low precision. Confidentiality is the exception, with F1 up to 95.7% in reasoning mode. Verdicts are unstable across runs: identical on 89.7% of goals for GPT reasoning, 74.0% for DeepSeek reasoning, 70.1% for GPT chat, and 61.6% for DeepSeek chat. Self-reported confidence is uniformly high yet shows no meaningful correlation with correctness. All results rest on a single zero-shot prompt and two model providers, which limits generalisability. On this benchmark, LLMs do not match formal verification, but may serve, at best, as pre-screening filters.

Electron is a popular framework for building cross-platform desktop applications using web technologies. Such applications consist of multiple processes with different privilege levels that communicate via message passing. When inter-process messages carry attacker-controlled inputs, they can propagate across processes and reach privileged APIs, e.g., command execution. Such a message propagation behavior is characterized as Message Progression Vulnerabilities (MPVs). The exploitation of MPVs is challenging because it often requires multiple steps, e.g., first arbitrary code execution in one process via message passing, and then command injection in another process using another message crafted in the first process. To our knowledge, existing works on Electron security only study unsafe configurations and malicious Document Object Model (DOM) content, i.e., they cannot detect or exploit these vulnerabilities that need to be triggered by complex cross-process exploits via message passing. We present Proton, a segmented directed fuzzing framework for detecting MPVs. Our key insight is to decompose end-to-end fuzzing into per-process segments along message-passing boundaries, where the goals of fuzzing each segment are either: (i) reaching a sink in the current process or (ii) propagating the payload to the next process, to enable the exploration of another process. In the second case, the messages seed the corpus of the next segment. Finally, Proton synthesizes crash inputs from each process to validate end-to-end exploits. We evaluate Proton against 589 real-world Electron applications, resulting in 23 zero-day MPVs. Among them, 22 lead to OS command execution, including projects with over 50k GitHub stars. We responsibly disclosed all findings. To date, we have received 13 acknowledgments, 11 fixes, and 11 CVEs, including a bug bounty from Vercel.

Building Automation Systems (BAS) manage critical building functions using protocols such as BACnet/IP, yet defenders have limited tooling and few labeled datasets for detecting BACnet-specific attacks. This work addresses these gaps through three contributions. First, CISA's Zeek BACnet parser is modified to produce a unified per-packet log, simplifying feature engineering for machine-learning (ML) pipelines. Second, a simulated BACnet/IP testbed is developed using bacpypes3 to model a small commercial HVAC system with physics-based device behavior, schedule-aware controller logic, and per-packet attack labeling. Third, five unsupervised anomaly detection models are evaluated using baseline traffic and six BACnet attack types, including denial of service, reconnaissance, property tampering, and false data injection. Results show that One-Class SVM achieved the strongest overall performance, with an average F1 score of 0.864 across all attacks and F1 scores above 0.99 for high-volume denial-of-service and reconnaissance attacks. Detection is much stronger for high-volume attacks, such as DoS attacks and reconnaissance, than stealthier techniques such as tampering and false data injection, which scored around 77%.

Perturbation techniques that turn unsuccessful jailbreak prompts into successful ones are continuously evolving, constituting a major security threat to LLM safety. In this paper, we investigate the internal representations of such string-level perturbed jailbreak inputs in the small weight models of the Qwen-2.5-1.5B/-3B/-7B-Instruct and Llama-3.2-1B/-3B/-3.1-8B-Instruct families. We select two representation spaces: the last-layer-last-token embedding space and the top-50 next-token probability space. The former space separates prompts based on their spelling and format, while the latter space is effectively one-dimensional but appears more complex to cluster. Within our refusal-dominated answer set we find no behavioral hyperplane in either space. Only the next token "Sure" in the 1.5B Qwen model, and both tokens "," and "ĊĊ" in the 1$ Llama model, display a significant association with a compliant-labeled answer.

Deepfake news refers to AI-generated (or AI ma-nipulated) multimedia content intentionally generated to deceive audiences by manipulating the facial expressions, or speech while maintaining the realistic appearance. The rapid progress of generative AI has made the synthesis of highly realistic fake videos and cloned voices widely accessible, posing a serious threat to the authenticity of digital news media. This paper presents a multi-modal framework that discerns the authenticity of video content by jointly exploiting audio and visual cues, thereby addressing the challenge of detecting the deepfake videos. We proposed a framework that involves features extraction from lip movements, audio content and video frames. Lip movements and speech content are encoded using the LipNet and DeepSpeech2 models, while facial features are extracted by leveraging the use of BlazeFace and represented with ResNet18. The extracted feature vectors are concatenated into a holistic video representation and classified with an ensemble of machine learning and deep learning models, including Random Forest (RF), Multi-layer Perceptron (MLP) and Long Short-Term Memory (LSTM) networks. Exten-sive experiments performed on the FakeAVCeleb dataset shows that the proposed approach attains an accuracy of 94% using augmented audio features, outperforming a state-of-the-art multi-modal ensemble baseline. The results confirm the robustness and practical potential of the proposed framework for deepfake news detection.

Broken Object-Level Authorization (BOLA), also known as Insecure Direct Object Reference (IDOR), has topped the OWASP API Security ranking since 2019 and is the root cause of some of the largest exposures of applicant data in recruitment technology. The defining feature of this flaw class is that a malicious request is byte-for-byte indistinguishable from a legitimate one, which is precisely why web application firewalls and single identity scanners fail to catch it. We present AuthProbe, an open-source, black-box scanner that detects BOLA and IDOR in HTTP APIs by driving its tests from an OpenAPI specification and by acting under two or more identities that the operator controls. AuthProbe discovers, for each identity, the objects that identity legitimately owns, then attempts to read one identity's objects while authenticated as another and confirms a leak by comparing the response against a ground-truth fetch by the true owner. It also walks predictable identifiers to expose enumeration and reports missing authentication and existence oracles. The tool returns a severity-thresholded exit code and machine-readable reports so that it can gate a continuous integration build. On a synthetic recruitment API in which the McHire failure class is reproduced, AuthProbe detects every planted cross-identity read with no false positives on a hardened counterpart, and its running time grows linearly with the number of objects under test. AuthProbe is released under the Apache 2.0 license with an authorized-use guardrail.

Nikolov and Ullman asked whether k statistical queries on a universe of size T can be released under pure differential privacy with expected worst-coordinate error at the square-root rate suggested by known lower bounds. We prove their conjectured upper bound. For every database size n and privacy parameter $\varepsilon>0$, there is an $\varepsilon$-differentially private mechanism with expected error $O(\min\{1,\sqrt{\log(2T)\log(2k)/(\varepsilon n)}\})$. This matches the lower-bound dependence in the standard high-dimensional regimes where those bounds apply; the shifted logarithms and outer minimum make the upper bound valid without additional parameter assumptions. The construction starts from a selection-only private multiplicative weights transcript, then replaces its probability mass function by a distance-penalized likelihood envelope. To prove that the modification preserves accuracy, a likelihood-level Maurey argument upper-bounds each Hamming-ball maximum by a small family of auxiliary PMW laws. Renyi moment bounds control nearby balls, a direct mixture bound controls distant balls, and grouping radii at the privacy scale prevents an additional $1/\varepsilon$ factor in the error. The mechanism is information-theoretic. A companion Lean 4 development machine-checks the finite construction, pure privacy after deterministic decoding, and the displayed all-regimes upper bound.

Gabidulin codes are a rank metric analog of Reed-Solomon codes. They are of great importance in rank-metric cryptography, as they are employed in highly efficient rank-based schemes, such as RQC and the Loidreau cryptosystem. However, until now, there has been no constant-time implementation of these codes, even though such an implementation is crucial for real-life development of cryptosystems. In this paper, we propose the first constant-time decoding algorithm of augmented Gabidulin (AG) codes, a simple variation of Gabidulin codes, in which columns of zeros are added to the end of a Gabidulin code. AG codes are used in the most efficient variants of the RQC cryptosystem. We prove that AG codes decoding can be achieved with quadratic complexity by reducing the decoding of an AG code to the decoding of a shorter standard Gabidulin code. In addition to the description of the complete AG codes decoding procedure, we present a constant-time algorithm for the left division of $q$-polynomials that can be used in every constant-time implementation of Gabidulin code variants. These algorithms are integrated into the RQC-Block-MS-AG scheme, and we evaluate the performance of our implementation through benchmarks. Our results show that our constant-time implementation outperforms the original RQC, though it remains approximately four to five times slower than HQC. However, it achieves ciphertext and key sizes about four times smaller, highlighting an appealing trade-off between performance and compactness.

Edge Intelligence (EI) has emerged as a transformative model for mission-critical unmanned platforms, such as drone swarms, by enabling collaborative model training at the network periphery. However, the security of FL deployments depends on both network availability and robust client authentication mechanisms. This paper investigates a chained attack against drone-based FL systems that combines network-layer denial-of-service with credential-based impersonation. We demonstrate that an adversary can: (1) force legitimate drones offline using 802.11 deauthentication attacks, and (2) subsequently impersonate the disconnected drone using extracted credentials. Through a systematic literature review and empirical validation using the Flower framework on two distinct testbeds of Raspberry Pi and Jetsons, we quantify the impact of availability disruptions under Independent and Identically Distributed (IID) and Non-Independently and Identically Distributed (Non-IID) data distributions, and confirm that single-factor authentication permits post-disconnect impersonation. Our findings reveal that even short-term wireless interruptions cascade into substantial training instability, particularly under non-IID conditions, while the authentication gap enables adversaries to seamlessly replace disconnected nodes. We discuss the compounded implications for mission-critical drone deployments and outline directions for future defenses addressing both availability and authentication vulnerabilities.

LLM-driven autonomous agents are reshaping offensive security. Unlike traditional penetration-testing tooling - deterministic, narrowly scoped, and operated by trained practitioners - agentic security tools exhibit indeterminacy along three independent dimensions. First, their actions are drawn from a non-deterministic policy whose outputs resist both ex-ante and ex-post explanation. This complicates incident attribution and pre-deployment safety reviews. Second, their impact is open-ended due to their non-deterministic actions, agency of utilized models, and opaque LLM supply-chains. Third, their user population is indeterminate in both size and required skill: the operating skill floor for using or developing offensive capabilities has dropped sharply. These three properties are linked thematically, but are not derivable from one another. Combined with the structural cost asymmetry between offense and defense, they enable the industrialization of offensive capability. The net short-term effect favors attackers, even if the same technology may, in the long run, democratize access to defensive practice. Existing dual-use cybersecurity and AI-ethics frameworks struggle to address this combination. Our work analyzes how moral attribution becomes diffuse between users, tool-makers, and third parties when employing autonomous AI agents for offensive security. We also examine the stakeholder impact of this technology and provide stratified recommendations.

Malware analysis demands rapid interpretation of complex detonation reports spanning filesystem, network, and process behaviours. While large language models (LLMs) demonstrate impressive capabilities for technical artifact interpretation, the opacity and escalating API costs of closed-weight frontier models motivate exploration of open-weight alternatives. However, many open-weight models are large, demanding significant compute resources and incurring non-trivial hosting costs that place them beyond reach for resource-constrained deployments. This paper investigates whether orchestrated ensembles of small language models (SLMs) can match or exceed single LLM performance on structured questions about malware detonation reports. We established baselines by testing eleven open-weight SLMs, three cyber security pre-trained models, and six frontier LLMs on Meta's CyberSecEval Malware Analysis benchmark. We then designed and evaluated four orchestration architectures: (i) a multi-agent pipeline that decomposes analysis into structured evidence-collection and reasoning stages, (ii) an adversarial debate framework in which two agents iteratively critique each other's reasoning, (iii) a hierarchical consultation system that pairs a general-purpose SLM with a cyber-specialised expert model, and (iv) a hybrid architecture that combines evidence-grounded pipelines with adversarial debate reasoning. The hybrid system (Qwen3-4B with Foundation-Sec-8B) achieved 35.30% overall accuracy, exceeding the strongest cyber-specialised baseline (22.54%) and the strongest ungrounded frontier baseline (34.77%); when given the same evidence pipeline, grounded Gemini remained the strongest configuration at 38.22%. These findings show that evidence-grounded orchestration can substantially improve the performance of collaborative SLMs for supporting interpretation of malware detonation reports.

Integrated sensing and communication (ISAC) enables the acquisition of environmental information by leveraging wireless signals transmitted for communication purposes. In this paper, we utilize this capability to reconstruct the layout of objects surrounding multiple receivers. Ray tracing is then applied to the reconstructed environment to infer the propagation channels for various transmitter positions, thereby constructing a channel knowledge map (CKM). The CKM is then used to verify the position of a legitimate transmitter, authenticating it against an adversarial device attempting to impersonate it from a different location. This physical layer authentication (PLA) mechanism utilizes the approximate known position of the legitimate transmitter, obtained, for instance, from the network as in cross-layer authentication, to compare the channel estimated from the received signal with the corresponding CKM data. We evaluate the impact on the PLA performance of both ISAC-induced CKM reconstruction errors and receiver-side channel estimation noise, in terms of false alarm and missed detection probabilities. Finally, the proposed approach is validated using an ISAC dataset from the literature.

An increase in advanced Android malware requires the use of deep learning models, which can run on Android devices. But there is a trade-off between security and energy use, as strong detection models can drain the battery of devices fast. This work tests different Multi-Layer Perceptron (MLP) model configurations to balance malware detection performance and energy efficiency. In this work, we compared standard FP32 models with optimized INT8 quantized neural networks with different model depths using TUANDROMD and DREBIN datasets for both classification performance and energy consumption. The results show that INT8 quantization reduces model size by about 3.5 times with a decrease in energy consumption to 0.0189 mJ per inference, while maintaining more than 99.2% detection accuracy. We found that shallow quantized architectures, such as 3-layer and 4-layer QNNs, reduce energy costs by improving throughput and shortening the time of CPU operating in a high-power state. This work shows that efficient malware protection can be achieved on resource-constrained smartphones and provides a foundation for Green AI in mobile security.

Key-Value (KV) cache reduces inference latency in large language models (LLMs). Traditional prefix-based reuse has low cache hit rates across inference requests because it requires exact token and position matches. To improve efficiency, recent system optimizations introduce position-independent KV reuse, allowing KV cache to be reused whenever identical text chunks appear, regardless of their position in the sequence. We show this design introduces a new threat, KV Cache Hijacking. Since KV caches are retrieved by token match but encode the context in which they were originally computed, the KV tied to a benign-looking token chunk may encode an attacker-controlled prefix. When later reused in a victim query, this contaminated KV silently hijacks the model's behavior, even if no attacker-controlled text appears in the input. We introduce HIJACKKV, the first attack framework that systematically exploits this vulnerability, demonstrating its severity and practicality. HIJACKKV optimizes an attacker-controlled prefix, so that the KV computed for a subsequent common benign text encodes the attacker's goal, while the text remains unchanged for future cache hits. HIJACKKV achieves an average 94% success rate in a single attempt, remains effective under realistic constraints including low hit rates (10%) and frequent recomputation (50%), persists over multi-turn interactions, and transfers across models in black-box settings. We further provide design insights for building secure KV reuse systems.

Agent safety is moving from content moderation toward preventing operational failures before tool-using agents act. We propose Janus, a foresight-oriented framework for long-horizon agent safety that trains guards to anticipate delayed risks from partial trajectories. Janus synthesizes diverse agent trajectories via multi-agent simulation and learns a shared policy with two coupled tasks: an anticipation task that forecasts safety-relevant futures and an adjudication task that decides safety from both the observed prefix and anticipated future. The two tasks are jointly optimized with CoAA-RL, which rewards forecasts by their utility for downstream safety judgment. The resulting guard model, Vanguard, blocks unsafe actions before execution. Across four agent-safety benchmarks, Vanguard improves average protection by 15.9 percentage points over baseline guards while increasing benign task completion by 5.1 percentage points.

Large language models (LLMs) are vulnerable to backdoor attacks, where hidden triggers induce malicious outputs. Existing defenses generally fall into inference-time detection or training-time mitigation, but face two key limitations. First, they focus on fine-tuning-based backdoors (e.g., PEFT modules) and fail to address insidious model-editing attacks that bypass training pipelines. Second, they target simple classification settings and do not naturally extend to open-ended LLM generation and do not naturally extend to the open-ended generation characteristics of LLMs. Consequently, these methods focus on surface-level behavioral patterns while neglecting the deeper representational causes of malicious activations. This lack of mechanistic understanding forces defenses to depend on empirical heuristics, limiting their robustness, generality, and practical applicability in real-world LLM deployment. To bridge this gap, we introduce DeCNIP (Defense with Critical Neuron Isolation Pruning), which leverages representational analysis to identify and neutralize backdoors in a unified pipeline. Specifically, DeCNIP identifies trigger-like behaviors by optimizing a cross-entropy loss between harmful prompts with candidate tokens and benign inputs. This representational discovery exposes latent threats by uncovering mechanisms through which triggers hijack model weights. It then isolates Backdoor Critical Neurons (BCNs) and prunes them selectively to remove malicious influence while preserving model utility. Extensive evaluations on six open-source LLMs and two benchmark datasets demonstrate that DeCNIP achieves over 95% relative reduction in Attack Success Rate (ASR), outperforming seven state-of-the-art defenses with only 0.1% neuron intervention. Moreover, it maintains 97% of the model's performance on normal benchmarks, demonstrating its efficacy, robustness, and scalability.

Adversarial robustness is commonly evaluated with predefined attack ensembles, such as AutoAttack, at a single perturbation budget $\varepsilon$ and on a selective choice of perturbation norms. We argue this formulation is fundamentally limited. First, robustness–perturbation curves may intersect or decay at different rates across models, making single-$\varepsilon$ rankings unstable. Second, current ensembles provide no evidence of optimality, leaving an unknown gap to worst-case performance. Third, fixed attack configurations provide no systematic control over the trade-off between attack strength and evaluation cost. To address these limitations, we introduce a unified evaluation framework based on a comprehensive pool of minimum-norm attacks and robustness–perturbation curves across $\ell_0$, $\ell_1$, $\ell_2$ and $\ell_\infty$ norms. We define the attack frontier as the worst-case robustness estimate the attack pool produces against a model. We then formalize evaluation as a frontier-approximation problem, constructing minimum-norm attack ensembles, optimized subsets of the comprehensive pool, that approach the frontier under a controllable query budget, with larger budgets monotonically tightening the estimate. Furthermore, we define the defense frontier as the maximum robustness across the model set at each perturbation size. We finally propose the Defense Optimality Index to rank defenses by their gap to the defense frontier, providing a ranking without selecting a reference $\varepsilon$. On CIFAR-10 and ImageNet, our ensembles match or exceed AutoAttack on most defenses at every budget tier, at fixed and controllable query cost, offering practitioners a query-controlled, curve-based alternative to fixed-$\varepsilon$ evaluation.

Traditional pentesting uses reconnaissance at each step to uncover unseen weaknesses, build stronger attacks, and advance the objective; we argue that AI agents require the same treatment. We formalize agent reconnaissance by modeling the process and identifying the knowledge assets it seeks to extract: what they are, how they are used, and which agent weaknesses they exploit to give adversaries leverage in indirect prompt injection attacks. We instantiate these insights in Know Your Agent (KYA), a framework that automates black-box, reconnaissance-driven pentesting by probing agents, building target profiles, and using those profiles to craft stronger attacks. We evaluate KYA on agent-security benchmarks and a real-world coding agent, and release KYA, its benchmarks, and baseline implementations for reproducibility.

Most existing LLM safety evaluation and defense methods are static: jailbreak vulnerabilities are assessed with fixed attacks, and guardrails are trained on fixed malicious-prompt datasets. In practice, adversaries continually evolve and expand the attack space. We propose DARWIN, an evolutionary attack-defense framework that models jailbreaking as a continual process and updates guardrails through an attack-defense loop. DARWIN-Attack evolves its capabilities through strategy discovery, mutation, selection, and feedback-guided composition. It mines strategies from broad external sources, generates variants through self-reflection and genetic evolution, filters them by performance against aligned LLMs, and adaptively selects and composes strategies during attacks. Through continuous evolution, DARWIN-Attack achieves state-of-the-art attack success rates against frontier LLMs and guardrails, reaching nearly 100% on DeepSeek-V4-Pro, over 90% on GPT-5.5, and nearly 100% on YuFeng-XGuard, outperforming recent evolving frameworks such as LSA and MAGIC. Its continued evolution exposes new vulnerabilities, motivating timely defense updates. Accordingly, DARWIN-Guard performs online adversarial training on emerging samples generated by DARWIN-Attack and jointly learns from malicious and benign disguised queries to recognize underlying intent rather than superficial attack patterns. DARWIN-Guard achieves 95.0% average unsafe recall across multiple safety benchmarks, outperforming recent advanced guardrails such as YuFeng and Nemotron, while maintaining nearly 100% average benign pass rate on standard benign benchmarks and the best performance on over-refusal benchmarks. Our code and model are available at https://github.com/ZJU-LLM-Safety/DARWIN.

Cyber Threat Intelligence (CTI) reports richly describe real-world attack processes, but their unstructured narratives cannot be directly used for automated attack-path reasoning. Existing CTI extraction methods focus on indicators, entities, or TTP labels without modeling the execution conditions and resulting states of each attack step, so the extracted knowledge supports neither state matching nor reachability analysis across multi-stage attack chains. This paper proposes an automated framework that extracts reachable attack chains by modeling each attack step as an attack unit of preconditions, an attack behavior, and postconditions. A multi-stage pipeline assisted by large language models (LLMs) extracts attack behavior skeletons, recovers their preconditions and postconditions, normalizes them into predefined predicates, and repairs broken dependencies; the resulting units are compiled into Datalog-style rules for attack-goal reachability reasoning. On a dataset of 20 CTI reports containing 334 human-validated annotated steps, our framework achieves higher annotated-step coverage than representative CTI extraction systems in recovering attack behaviors. Moreover, by explicitly generating preconditions and postconditions, it produces attack units that are more complete and consistent than those generated by end-to-end LLM baselines. On the extracted chains, Datalog inference reaches the specified attack goal in 19 of 20 reports, while backward search yields 34 attack paths under the generated rules. The source code and experimental artifacts are available in an anonymized repository. .

Vision-Language Models (VLMs) are known to be vulnerable to adversarial attacks, where subtle perturbations to images or texts induce erroneous outputs. However, most text-based attacks are adapted from language-model-centric methods, in which the visual input is fixed during optimization, resulting in adversarial prompts that are tied to specific images and thus limiting their attack effectiveness. To this end, we first introduce a new research perspective: cross-image transferability for adversarial prompts. We then propose GhostPrompt, an adversarial prompt that is optimized once and reused to steer VLM outputs toward attacker-specified responses across diverse images. GhostPrompt employs a joint optimization that distills image-invariant adversarial features into the prompt by "worst-case" generation. Specifically, it alternates between constructing hard visual conditions for the current prompt and updating the prompt to remain effective under these conditions. Extensive experiments on prevalent VLMs verify that \ourmethod achieves an improvement of over 30% in attack success rates compared to state-of-the-art (SoTA) baselines, while reducing computation time by ~70%. Our code is avalable at https://github.com/Ye-ze-yu/GhostPrompt.

Federated Graph Neural Networks (FedGNNs) are highly vulnerable to backdoor poisoning, yet existing defenses typically rely on rule-based approaches that lack semantic understanding, making them vulnerable to stealthy triggers and harmful to benign structures. To solve this, we present FedLSG, the first framework that integrates large language models (LLMs) into federated graph backdoor defense. FedLSG introduces a graph and behavior to text grounding scheme that transforms local graph structures and client update behaviors into semantically rich natural language representations. The framework further adopts a lightweight student-teacher architecture. On the server side, a full scale LLM serves as a teacher, providing global contextual guidance and evaluating client updates during aggregation to identify potentially malicious participants. On the client side, a LoRA-based student is maintained to perform semantic reasoning, to suppress the influence of edges associated with backdoor triggers. By enabling semantic interpretation of both graph patterns and client behaviors, the framework adaptively incorporates rule-based signals into message passing and client aggregation for defense. Experiments demonstrate that FedLSG significantly improves resistance to backdoor attacks without compromising graph integrity.

Many administrative processes, such as internship agreement processes, often rely on manual approval workflows and centralized record-keeping. This makes the process susceptible to delays and unauthorized modification while introducing limited traceability. This study presents StajChain, a permissioned blockchain-based multi-party internship management system developed using Hyperledger Fabric. The proposed system implements the complete internship agreement lifecycle through smart contracts and enforces role-based authorization using Hyperledger Fabric CA. The architecture consists of a React frontend, a NodeJS backend, an off-chain SQLite database, and the on-chain Fabric ledger. Users such as students, companies, faculty internship committee members, and the central internship unit can perform specified operations according to their role and identity. The agreement lifecycle follows predefined sequential steps, and at each step, the ledger status is updated and recorded securely. Furthermore, the system was evaluated using functional and performance tests, indicating acceptable throughput and latency for verifiable administrative workflows. This implementation demonstrates how permissioned blockchain technology can improve transparency, integrity, and accountability while preserving controlled access to institutional data and providing a working prototype that can be used in various future systems.