Papers

Current approaches to LLM adversarial testing suffer from coverage gaps: manual red-teaming does not scale, LLM-as-attacker methods exhibit mode collapse, and gradient-based approaches produce uninterpretable gibberish. We introduce a quality-diversity evolutionary framework that operates at the semantic level, evolving interpretable attack strategies rather than token sequences. Using MAP-Elites, we maintain a diverse archive of attacks across behavioral dimensions (strategy type, encoding method, prompt length). In experiments across GPT-4o-mini, Claude 3.5 Sonnet, Gemini 2.0 Flash, and an open-weight coding model (Devstral-small-2), we discover distinct vulnerability profiles: GPT-4o-mini is vulnerable to hypothetical and multi-turn framing combined with ROT13 encoding (fitness 0.8), Gemini to direct attacks with ROT13 and multi-turn with Leetspeak (0.8), while Claude shows uniformly ambiguous responses across all strategies (max 0.4). The semantic representation produces interpretable attacks that reveal systematic, model-specific weaknesses, providing actionable insights for improving LLM safety and a reproducible baseline for evaluating future frontier models. Code and experiment artifacts are released at https://github.com/bassrehab/red-queen.

Vulnerability discovery on C/C++ source asks the analyst to choose between heavyweight static analysers, which need a working build before a single query runs, and free-form LLMs, which read source readily but invent details and lose track of cross-function dataflow on real codebases. We present NeuroLog, an end-to-end build-free pipeline that assigns each layer the role it is uniquely good at: an LLM extracts typed dataflow facts one function at a time; a Souffle rule mesh composes those facts into cross-function findings; a Z3 post-pass filters infeasible findings and emits a SAT model for each survivor. To go beyond pure static reasoning we also fold in runtime evidence: likely range invariants from a handful of corpus seeds tighten the SMT problem at near-zero cost. A second LLM agent reads each SAT model and writes a Python program that produces a candidate crashing input, validated by an AddressSanitizer harness. Combining static-narrowing-SMT (Saturn, Pinpoint) and Datalog-with-SMT (Formulog) is prior art; new here are an LLM-derived fact base, a no-build pipeline, and the SAT model as an artifact (input to crash synthesis) rather than a yes/no verdict. Across stb, cJSON, libxml2, an FFmpeg demuxer slice, and curl 8.3.0, NeuroLog re-discovers eight CVE-class issues end-to-end, including the CVSS-9.8 SOCKS5 heap overflow CVE-2023-38545, each ASan-confirmed. On libarchive HEAD we surface five memory-safety bugs (four previously unreported) across the cpio reader and the XAR/WARC/7zip writers; all filed upstream, several fixes merged, with the cpio use-after-free acknowledged in seven hours. Extraction takes ~37 s and $0.005 on stb; crash synthesis turned a static finding into a 102-byte stb_vorbis crash in two LLM iterations (no fuzzer); a likely-invariant filter from three Matroska seeds eliminates 13.2% of the FFmpeg-demuxer feasible set.

Current backdoor attacks against LLMs are typically manipulated by the attacker and remain passive. In this paper, we introduce the Proactive Availability Backdoor (PAB), a novel paradigm that shifts the attack vector from passive waiting to active social engineering. By weaponizing the inherent helpfulness of aligned LLMs, PAB proactively traps users into executing trigger-implanted queries by offering suggestions, achieving high aggressiveness, precision and stealthiness. To rigorously evaluate its threat in a real-life context, we introduce a dual-agent ecological simulation framework based on selected dimensions of the Five-Factor Model, and deploy PAB with few-shot prompts. Being validated on different models and domains, PAB performs remarkably and its effective attack success rate, which calculates the joint probability of attack incidence rate and attack success rate, goes to 73.1%. We also introduce Anti-PAB, a defense method tailored for PAB. Our findings reveal that the helpfulness of LLMs can be weaponized to compromise availability, exposing a serious hidden threat to LLMs users. We release all the scripts and datasets in the experiments at https://anonymous.4open.science/r/PAB-ANONYMOUS/.

Reasoning traces have become a valuable form of learning signals for improving and transferring the capabilities of large language models. In particular, detailed traces can help distill reasoning behavior from stronger teacher models into weaker student models. The value of capability transfer has motivated many deployed systems with reasoning models to hide raw internal traces and expose at most summaries and answers to users. As a result, we ask whether such interface-level trace hiding prevents users from obtaining useful reasoning supervision through prompting. We study this question with Reasoning Exposure Prompting (REP), a lightweight in-context elicitation method that uses shadow-model-generated demonstrations wrapped in auxiliary code-like formats to raise user-visible reasoning traces from a victim model. Across the common reasoning dataset, different victim models, and different student model distillation, REP substantially increases similarity between exposed and REP-conditioned internal traces while preserving useful reasoning signals.

Reproducing software vulnerabilities is fundamental to security researchers, open-source maintainers, and educators. Yet, vulnerabilities remain hard to reproduce today, and even when they can be reproduced, recreating a software environment where the vulnerability can be exploited becomes harder and harder over time. We present NICE, the NIx CvE reproduction framework, which uses declarative recipes to build and automatically validate vulnerable environments. In NICE, a reproduced CVE comprises one or more NixOS virtual machine configurations, a scripted exploitation scenario, and machine-checkable assertions that provide factual evidence of exploitation. This design facilitates sharing, validation, review, and long-term reproducibility. We evaluate NICE on 19 diverse real-world CVEs spanning multiple CWE categories, attack vectors, and target types (user-space, system software, kernel, and graphical applications). We show that NICE allows to produce concise recipes and integration tests that reproduce vulnerable environments and provide proofs of exploitation. NICE is applicable to security education and training (e.g., creating cyber ranges), but also to vulnerability reporting, where its reproducibility and reviewability properties can make reports easier to audit and verify.

Generative artificial intelligence has fundamentally changed how content is now produced. It has enabled how high-fidelity text, images, audio, and videos are created, modified, and redistributed at near-zero marginal cost. This shift exposes enterprises and ecosystems to a number of risks across four reinforcing authenticity layers – authenticity, provenance, integrity, and accountability – that traditional controls are inadequate to address in isolation. We introduce the concept of authenticity debt: the cumulative institutional liability that accumulates when organizations deploy AI-generated content without preserving verifiable origin, integrity, and accountability, deferring exposure that surfaces under regulatory, legal, or market scrutiny. This paper presents a comprehensive, multi-dimensional taxonomy of generative AI harms and attack vectors, surveys the capabilities and failure modes of technical controls including digital watermarking, provenance frameworks (C2PA, Adobe CAI), and detection technologies, and argues that no single mechanism is sufficient in open, adversarial, and evolving environments. Drawing on Zero Trust Architecture principles and enterprise governance frameworks, we propose a layered reference architecture that integrates cryptographic provenance, human-in-the-loop verification, and continuous governance to sustain defensible authenticity at scale. We further examine the regulatory landscape (EU AI Act, U.S.\ FTC, NIST AI RMF) and identify practical guiding principles for organizations seeking to build authenticity as institutional infrastructure rather than an afterthought.

As language models take on agentic roles that call APIs, read tool outputs, and act on third-party content, their attack surface expands beyond what users type. Whether they treat a malicious instruction the same way regardless of where it arrives has not been studied systematically. We introduce the Safety Asymmetry Score (SAS), measuring how a model's susceptibility to adversarial content shifts depending on whether it arrives in the user message, tool metadata, or tool output, using matched payload pairs that hold the malicious text identical and vary only the channel. Across 10 production LLMs and three attack families, general-purpose models sharply discount instructions arriving as tool metadata relative to identical instructions in the user message, while agent-native models discount them far less. This differential survives an affordance-matched control equalizing tool availability and scoring, and a size-controlled mixed-effects analysis. Within the tool surface the pattern reverses: the same content is far more instruction-like in a tool's description than in its output, with affordances identical. Models treat tool metadata as instructions and tool results as data. In Llama 3.3-70B and GPT-OSS-120B this signal is causally present at mid-to-late depths but non-linearly encoded, so linear probes miss what activation patching recovers.

Deceptive web content, widely instantiated across the internet and commonly known as social-engineering attacks, manipulates autonomous web agents into submitting users' personally identifiable information (PII) to attacker-controlled endpoints. In this paper, we show that social-engineering attacks are highly effective at extracting critical-tier PII from frontier web agents, posing a severe risk to deployed agentic systems. To quantify this risk, we introduce \textbf{Scammer4U}, a pre-registered benchmark of 91 attacker-controlled environments and 10 benign-twin baselines, spanning 8 attack vectors and 16 site categories on an 8-axis factorial taxonomy that isolates the causal contribution of individual attack design factors. Across frontier agents, we find that critical-tier PII leakage reaches 54–93% under no privacy guidance, compared to 0% on benign-twin baselines, confirming that leakage is attack-attributable rather than incidental form-filling. Escalating prompt-level mitigation yields sharply model-dependent reductions across the four families and remains insufficient to reliably prevent critical PII submission at the pooled level. Most critically, we identify a detection–action gap: agents whose reasoning an independent LLM judge confirms has flagged the site as suspicious still submit critical PII in 35.9% of sessions, versus 66.1% when no suspicion is verbalized, a 30.2% gap robust across all four model families. Our findings reveal that defenses conditioned on the agent's own recognition of an attack are gating on the wrong signal, motivating output-level interception of outbound submissions that operates independently of the agent's reasoning loop.

ChatGPT Apps, launched by OpenAI on Oct. 6, 2025, introduce an app-in-app paradigm in which third-party applications share a single chat context with the user and with every other connected app. The ecosystem grew from 122 apps in Dec. 2025 to 888 by May 2026, yet its security has remained uninvestigated. We identify cross-app context poisoning, a variant of indirect prompt injection distinguished by three properties: 1) the injection persists in the shared chat context across turns; 2) the effect surfaces through a different co-resident app the user later invokes; and 3) the delivery vectors are first-party APIs exposed to every connected app. We find multiple APIs capable of writing app-controlled content into the shared context, with sendFollowUpMessage as the most direct and potent channel. Two undocumented parameters that the runtime silently accepts, systemPrompt and isVisible, amplify this channel to silent, system-priority writes. Leveraging this channel, we realize a confused-deputy attack in which a malicious app poisons the context so that the LLM, consulting that context, enables manipulation against benign co-resident apps. We demonstrate two payload styles (conditional and imperative) and evaluate them across six current ChatGPT models. The root cause is architectural: the LLM's context is a persistent, flat, untagged data store shared by user and apps, with no isolation. Every mature multi-tenant platform, from Multics virtual memory to Android UIDs and iOS sandbox profiles, paid the isolation cost before admitting third parties; ChatGPT Apps did not. Fixing this requires an architectural change, not a patch. We disclosed our findings to OpenAI; the undocumented parameters remain accessible at the time of writing, and the architectural gap is by design: the shared context that enables cross-app composition is the same flat namespace that enables cross-app poisoning.

This research presents a novel stochastic framework for proactive cybersecurity defense timing under a single attack scenario. The approach models the defense process as a continuous observation mechanism in which the defense instant and the subsequent observation slot follow independent exponential distributions. Laplace-Carson transforms combined with first-excess theory yield the joint detection function that brackets the attack moment. Marginalization under Markovian Poisson arrivals then produces the probability density of the defense moment and conditional expectations of pre-attack and post-attack observation times. These closed-form results enable quantitative assessment of defense timing sensitivity to threat intensity and support precise calibration of observation parameters for low-latency proactive measures. Major contributions include the explicit derivation of marginal distributions and expected values, visualization of defense moment density, and the bridging of stochastic duel methodology with practical cybersecurity applications.

Coverage-guided kernel fuzzers such as syzkaller rely on edge coverage trace-pc as their sole feedback signal. This context-blind approach cannot distinguish execution paths that differ only in argument values. For example, two invocations of copy_from_user() with different size parameters hit identical basic blocks yet have vastly different security implications. I present, an LLVM-based instrumentation framework that extends Linux KCOV with data-flow extraction of function arguments and return values. A compiler pass emits lightweight callbacks that report one source-level parameter per call, passing the value itself in a register together with its index and byte size. A composite is instead reported by the address of the object holding it, paired with a constant table of field (offset, size) pairs derived from DWARF DICompositeType metadata, with zero source annotation. The callbacks carry no program counter. The kernel derives it from the return address, which also attributes records the compiler emitted from inside an inlined body to the callee that was inlined. A lock-free per-task record buffer delivers variable-length records to user space with no interference to existing KCOV or syzkaller infrastructure, carrying entry, return, and comparison-operand records in one unified stream. I demonstrate dual utility: (1) fuzzers gain state-aware feedback for mutation guidance into value-dependent state transitions, and (2) security analysts obtain deterministic argument records for root-cause analysis without printk or kprobe overhead.

LLM agents increasingly rely on community-contributed skills that expand an agent's operational capability set. We study a core safety problem in agentic AI systems: whether individually safe skills can compose into unsafe installed skill sets. We present SkillReact, a compositional security measurement framework with three components: a deterministic static-composition benchmark, a two-rater LLM-assisted human-adjudication pipeline, and an action-based exploitability harness. On 1,520 ClawHub skills, 651 pass individual inspection and form 211,575 pairs; the benchmark flags 22.25% of these as structural candidates. We treat this raw rate as a recall-oriented scanner ceiling and calibrate it against human judgment: in a pattern-stratified audit, roughly one in five flagged pair-pattern hits survives as a real compositional risk (population-weighted validity 18.2%, our headline result), implying about 14K genuine risk memberships in a single registry that per-skill scanning misses by construction, since every pair is individually safe. An action-based harness then probes when these candidates become model-issued tool calls, and finds realization gated by host-model disposition: on an anchor-conditioned dropper subset, Haiku-4-5 issues the dropper-stage tool call on all 39 direct-prompt trials (36 of them the full download-then-execute chain, 3 download-only), Opus-4-7 stops at the download, and Sonnet-4-6 refuses outright. A control that holds the request fixed and varies only the installed skills finds compliance highest with no skills installed: a composition fixes which capabilities are reachable, while the host model decides whether to use them. Together these motivate install-time compositional checks and capability isolation as complements to per-skill scanning.

We study the problem of differentially private (DP) $k$-means clustering in Euclidean space. Previous solutions rely on summing the private data directly, which induces a sensitivity proportional to the domain. We introduce PE-means, an extension of the private evolution (PE) algorithm (an increasingly popular method for synthetic data generation), to the problem of $k$-means clustering. The key advantage of PE is that it only computes a private histogram with constant sensitivity to guide the evolution. Our adaptation of PE includes new evolutionary operators for clustering, as well as other algorithmic improvements of independent interest. Overall, PE-means achieves an average improvement of 26% in clustering loss over state-of-the-art baselines such as Google's LSH-based algorithm and DP-Lloyd variants.

Verifying claims about AI workloads is a prerequisite for credible AI governance of covert adversaries (who comply with monitoring only when detection likelihood is high), yet the apparent non-determinism of GPU floating-point arithmetic forces auditors to accept approximate output matches. Covert adversaries can exploit unverifiable degrees of freedom in monitored computation. Attack vectors include steganography, unreported modification of inference software, and covert computation via unreported batch elements. Empirically, we analyze how modern inference engines (vLLM, HF transformers) produce deterministic but non-invariant outputs, without needing to set performance-compromising determinism flags, if the right information is available for re-computation and no atomic functions are called in the backend. We demonstrate that such bitwise-precise re-computation does not require access to identical hardware, via a software-only emulation of LLM inference across multiple NVIDIA GPU variants. Thus, accumulated rounding errors can be an auditable signature of the software and hardware setup used for inference, instead of a constraint on verifiability.

Social deduction games, or hidden-role games, are multiplayer games in which players are assigned private roles and act under asymmetric information about other players' roles and actions. In the canonical example Werewolf, werewolves conceal their roles and mislead the other players, while the seer can obtain role information about a chosen player. Thus, a central functionality of such games is controlling which players can access which information. In typical play, this control is implemented by a trusted human moderator, who assigns roles, mediates secret actions, and reveals outcomes. This reliance raises the barrier to participation and introduces a trusted third party as a single point of failure. In this work, we show that Werewolf can be played without a moderator or any digital device, using only ordinary playing cards. Our construction maintains a shared pool of cards that is observable to all players and manipulated according to a common public procedure, while its interpretation depends on each player's private role. This induces role-dependent views from a single public sequence of card operations. Consequently, even without private messages, werewolves can identify one another and coordinate, and the seer can test whether a chosen player is a werewolf in each round. The proposed implementation is built from card-based physical cryptographic primitives, such as face-down commitments and verifiable shuffles, and higher-level subprotocols for intra-role information sharing, secret action designation, and attribute testing. These subprotocols implement the moderator's core functions while keeping all card operations public and auditable under standard assumptions on physical card operations. We show that the resulting complete moderatorless implementation of Werewolf scales to an arbitrary number n of players using O(n^3) cards.

Large language models (LLMs) are rapidly transforming how software is created and maintained. Comparing LLM-generated code against human-written standards is essential to determine whether these new tools uphold or erode the security baselines established by professional developers. Yet, we lack a standardized method for empirically comparing the security of code produced through human-LLM collaboration against LLM-only, or traditional human-only methods. To facilitate this, we propose an automated framework for conducting comparative studies across human-only, LLM-only, and hybrid conditions. Our approach automates the logging of prompts, timing, and experimental settings, measuring outcomes through multi-dimensional static and dynamic quality analysis. We provide an open-source implementation of this framework to ensure that future researchers can conduct reproducible, species-fair experiments. Importantly, we validate the framework via a feasibility study, providing an experimental blueprint for “species-fair” comparisons between human and AI subjects. By sharing lessons learned, we establish a foundation for empirical research on human and LLM-generated code for software security.

Mobile ad hoc networks (MANETs) based on proactive routing protocols such as OLSR, remain vulnerable to routing-layer attacks. While prior work has focused primarily on attack detection, the problem of identifying deployed defenses has received comparatively little attention. This work examines whether a routing-layer defense leaves detectable signatures in network traffic. The evaluated fictive mitigation mechanism operates entirely within standard OLSR control traffic and introduces no new packet types, making passive detection inherently difficult. Using ns-3 simulations across baseline, attack-only, defense-only, and combined attack-defense regimes under both static and mobile conditions, we derive features from observable routing dynamics and control-plane activity available to a passive attacker. Despite the restricted observability available to the adversary, the results show that defense detection remains feasible in this setting. Ensemble models achieve in-domain accuracy up to $0.91$ (AUC $0.96$). Cross-domain generalization is asymmetric: models trained on static data degrade under mobility ($\approx 0.67$), whereas mobile-trained models transfer more robustly ($\approx 0.84$). Restricting the model to a compact invariant feature subset of four metrics yields near-symmetric cross-domain transfer ($\approx 0.86$ in both directions). Further analysis shows that the cross-domain gap reflects both reduced class separability and decision-threshold transfer, with the latter largely recoverable through limited target-domain calibration. These findings indicate that the evaluated defense mechanism leaves a detectable statistical footprint in passively observable routing behavior, providing adversaries with a potential reconnaissance capability in protected MANET deployments.

The detection of intrusions in IoT-based networks poses challenges that cannot be overcome using traditional machine learning methods. Perhaps the biggest of them is related to the presence of a class imbalance in the side-channel dataset, where the number of samples in the normal class compared to the attacks can reach a ratio of 75,964 to 1. Such an aspect is addressed by Dominguez et al. through the proof of concept of power-based intrusion detection. Unfortunately, neither the authors attempt to cope with the problem of imbalance nor do they assess the classifier performance using a balanced training set. In the current paper, both aspects will be handled at once. First, a Synthetic Minority Oversampling Technique (SMOTE) was performed on all nine possible datasets extracted from the initial one, providing an exact imbalance ratio of 1.1 for each. Then, eight algorithms i.e. Random Forest, HistGradientBoosting, LightGBM, Extra Trees, XGBoost, k-Nearest Neighbors, Multi-Layer Perceptron, and Decision Tree were trained under identical conditions for the SMOTE balanced 6-hour dataset. Random Forest reached a micro-averaged F1 score of 0.9989 and macro F1 of 0.9794, thus outperforming the previously best micro-F1 result obtained by Time Series Forest algorithm from the base paper of 0.9983. Extra Trees provided the same performance as well, but at 10 times faster. The introduction of a macro-F1 metric explicitly in contrast to the base paper assessment reveals important class-level information missed with aggregate performance metrics. Recall rates per-class calculated with confusion matrices, F1 heatmaps, and ROC curves show that minority attack classes, especially those with combined M+L infections, are detected reliably only when using SMOTE balance. Feature importance analysis indicates the latest time steps as the most important predictor signals out of 60 steps in a power window.

Large language models (LLMs) suffer from degraded safety capabilities even when fine-tuned with benign datasets. However, existing methods for identifying safety-degrading samples in benign datasets suffer from high computational costs and significant noise issues. In this paper, we propose DataShield to efficiently and effectively identify potential safety-degrading samples. Our key intuition is based on the observation that benign fine-tuning increases the overall response compliance of LLMs. DataShield's key technical insight is to quantify each sample's contribution to the model's compliance behavior as its safety degradation score. DataShield consists of three core components: (1) Compliance Vector Extraction, which captures the LLM's compliance behavior tendency; (2) a novel Compliance-Aware Score (CAS), which automatically identifies the optimal safety-critical layer; and (3) Safety-degrading Sample Filtering, which quantifies the projection shift of training data along the compliance direction. Extensive experimental evaluation on Llama3-8B, Llama3.1-8B, and Qwen2.5-7B using the Alpaca and Dolly benign datasets validates our method's effectiveness in identifying high-risk and low-risk data subsets. We also observe that open-ended question answering is more likely to trigger safety degradation, and corresponding responses tend to be longer. We hope this work can provide new insights into data-centric defense methods. The source code is available at: https://github.com/ZJunBo/DataShield.

Modern network intrusion detection systems (NIDS) are caught in a structural contradiction: the protocols carrying the highest threat intelligence are precisely those encrypted under TLS 1.3 and QUIC, where payload inspection yields nothing. We ask a simpler question – what if the attack signature is not in the bytes, but in the rhythm? – and answer it by treating network flows as a language whose grammar is written entirely in L3/L4 packet metadata: length, inter-arrival time, TTL, TCP flags, and hashed port numbers. We present PLM-NIDS, which proves three claims in sequence. (1) The grammar exists and is learnable: a RWKV-4 state-space model trained on 344,232 unlabelled Monday flows achieves a causal LM validation loss of 0.204, demonstrating that benign traffic has predictable, statistically consistent structure. (2) Attacks violate this grammar: the per-flow perplexity score cleanly separates benign from attack flows with PR-AUC = 0.93 using zero attack labels at training time. (3) This separation is architecturally nontrivial: an LSTM trained on identical token sequences degenerates to a majority-class predictor (ROC-AUC approximately 0.50, F1 = 0.91 by always predicting "attack"), proving that RWKV's causal pre-training provides an inductive bias unavailable to direct classifiers. Supervised fine-tuning further raises PR-AUC to 0.94 and ROC-AUC to 0.75, with a precision of 97.7% at the calibrated operating threshold. The RWKV backbone's O(T) recurrent inference enables per-packet streaming without flow buffering, making PLM-NIDS operationally viable at line rate. Because it reads only IP/TCP/UDP headers, it is inherently encryption-agnostic: TLS 1.3, QUIC, and future encrypted protocols are handled transparently.

LLM-based agents are rapidly advancing, autonomously invoking external tools to complete multi-step tasks for users. However, agents often acquire more sensitive information than the task requires. Existing privacy benchmarks audit what the agent's response or outgoing actions disclose, but overlook the acquisition stage where data first enters the agent's context. The over-acquired information is then one careless action or one attack away from an outright leak. To assess its prevalence, we introduce PrivacyPeek, a benchmark for evaluating acquisition-stage privacy leakage of LLM-based agents, with $1{,}182$ cases across $7$ acquisition behaviours and $16$ application domains. Specifically, Acquisition Inspection examines the agent's tool-call trajectory, both the tools it invokes and the data it receives, to detect when it acquires sensitive information beyond the task scope. Probe Elicitation then issues a follow-up probe and measures how readily an attacker could elicit sensitive information the agent acquired but did not disclose. Our experiments on $10$ LLM-based agents across $4$ model families show that the unnecessary acquisition of sensitive information is widespread. In addition, we observe a correlation between the task-completion capability and acquisition-stage leakage. Prompt-level defences reduce only a small fraction of acquisition-stage leakage, leaving the majority unmitigated. These results make auditing acquisition-stage privacy both urgent and necessary. Our dataset and code are available at https://github.com/Xuan269/PrivacyPeek-Resource.

As Large Language Models evolve for user convenience, vulnerability to jailbreak attacks continues to be reported despite ongoing efforts in safety training. Traditional jailbreak techniques typically focus on a single prompt injection, neglecting the models' ability to remember the flow of conversation and the user's instructions. In this paper, we propose Persona Attack, a memory injection based jailbreak method that manipulates the model's context window through a step by step approach. Experimental results from applying Persona Attack to several widely used LLMs reveal that, as injections accumulate in memory, models increasingly prioritize these instructions over their internal safety alignment mechanisms. Furthermore, our experiments empirically demonstrate that the attack success rate varies not only according to the memory implementation of the model, but also combinations of instructions and can reach 95% under specific instruction configurations.

Language models can find thousands of severe software vulnerabilities, and agents are increasingly being misused for cyberattacks. To avoid detection, attackers frequently distribute their misuse, splitting a harmful task across many user accounts so each individual transcript looks benign. Because safety monitors score only one agent context at a time, they are structurally blind to misuse that is only visible in aggregate, across many accounts. We show this gap is real by building, to our knowledge, the first distributed agent attack, a multi-agent scaffold that completes hard cybersecurity tasks while hiding the harmful objective across subagents with limited contexts, evading a standard monitor that catches it only a fifth as often as prior agent attacks. Towards a defense, we develop an online stateful monitor that uses real-time clustering to collect weak suspiciousness signals across many agent transcripts, and escalates only rarely to a language model that flags misuse across user accounts. In evaluations with large-scale simulated datacenter traffic, our monitor Pareto dominates standard monitors, catching distributed attacks 30% earlier and flagging cyber misuse before it reaches the most harmful stages. Crucially, this comes at negligible additional latency for ~99% of user traffic. This detection advantage persists but narrows as the benign background traffic grows very large. After an extensive red-teaming exercise, we improve the defense and surprisingly also find that it catches standard jailbreaks, since adaptive attackers reuse attack variants across accounts. Our results point toward a new class of safety monitors which reason over groups of users rather than isolated transcripts.

Credential leakage in public source code repositories poses a critical security threat, with over 23.8 million secrets exposed in 2024 alone. Existing detection tools suffer from high false-positive rates because rigid pattern matching and binary classification schemes fail to distinguish genuine credentials from placeholder or weak credentials. We propose a three-class classification framework that explicitly models placeholder or weak credentials as a distinct class, leveraging CodeBERT-based semantic understanding combined with character-level pattern recognition. We evaluate our approach on a newly constructed dataset of 9,426 samples spanning 10 programming languages. Our model achieves a Matthews Correlation Coefficient of 0.86 and a macro F1-score of 0.90, achieving 93% recall and 89% precision for genuine credential leaks while reducing high severity alerts by 33.0% (from 373 to 250) without sacrificing security coverage. Compared to prior character-level approaches, our method improves placeholder or weak credential detection from 54% to 81% F1-score while maintaining strong cross language generalization, with 9 of 10 languages achieving F1 above 0.80 under leave-one-language-out evaluation.

We construct a family of 2D-local constant-depth quantum circuits that output states whose entanglement entropy across a specified cut cannot be estimated in quantum polynomial time. As constant-depth quantum circuits can be learned from polynomially many quantum samples, our resulting pseudoentangled states are implicitly public-key and not pseudorandom. This separates pseudoentanglement from pseudorandomness in the shallow-circuit regime: the former is possible, while the latter is not. The construction is based on the quantum intractability of the Dense-Sparse Learning Parity with Noise problem introduced in [DJ25] and uses a bounded-fan-in, bounded-fan-out classical randomized encoding for linear maps $\mathbf{x} \mapsto \mathbf{Mx},$ which could be of independent interest. As applications, we obtain quantum hardness for the problem of learning the entanglement structure (across a fixed cut) of the ground-state of 1D and 2D local Hamiltonians. The 1D Hamiltonian has an inverse polynomial gap, whereas the 2D one has a constant gap. This complements the result of [BZZ24] that showed only factoring-based hardness for the 1D case, though achieving a volume versus area entanglement difference.

Recent large language models (LLMs) incorporate reasoning capabilities that allow them to perform well in predicting whether a smart contract respects a certain property, suggesting a complementary approach to traditional formal-methods-based techniques for smart contract verification. However, the application of LLMs in such context has two major issues: 1) properties expressed in natural language are intrinsically ambiguous, and 2) answers returned by LLMs have no guarantee of correctness. In this paper, we address both issues simultaneously by: 1) introducing a new formal specification language that extends Solidity with abstract types, and 2) designing a workflow that combines LLMs with type checking and concrete execution to generate and validate violation witnesses (i.e., counterexamples). The key idea is to represent a specification as a Solidity test with (existentially quantified) variables of abstract type; finding an instantiation of these variables to concrete values (of the correct type) concretizes the test into an executable counterexample (PoC) for the target property. We implemented our procedure in the tool Neuroforger, experimentally evaluating it on a smart-contract verification dataset drawn from literature, obtaining promising results that demonstrate its potential applicability in the wild.

Financial inclusion has expanded significantly across Africa through mobile money services delivered primarily via USSD technology. However, visually impaired individuals continue to face accessibility and security barriers when conducting financial transactions. Current USSD systems are not designed for non-visual interaction, forcing users to rely on third-party assistance even for PIN entry, thereby increasing fraud exposure and reducing transaction confidence. Although alternative assistive technologies such as screen readers exist, they are not compatible with USSD operations, often causing sessions to time out before the user can complete a transaction. This paper presents an Android-based intelligent middleware that automates USSD transactions, integrates biometric-secured PIN injection, and introduces a privacy-preserving screen-dimming mechanism: Blackout Mode. The system leverages Android Accessibility Services, hardware-backed Keystore security, and on-device natural language parsing to enable independent, secure voice-based mobile money access. We show that the proposed solution improves task success rates from 65-75% to more than 90% and reduces transaction completion time from 40-60 seconds to 12-15 seconds, while also improving perceived security.

Reliable identification of encrypted data fragments is essential in cybersecurity, with applications to ransomware detection, digital forensics, and large-scale data analysis. Distinguishing encrypted from compressed fragments is particularly challenging, as short fragments lack structural data and exhibit low statistical redundancy. Traditional statistical methods based on byte-level distributions show limited effectiveness on this task. Recent machine learning approaches improve performance by learning subtle patterns from raw bytes, but predominantly rely on single-modal representations, implicitly assuming that a single view of the data is sufficient for accurate classification. This paper shows that this assumption becomes a fundamental limitation in low-information settings, when only small fragments of data are available (512–2048 Bytes). We propose Triumvir, a multi-modal, uncertainty-aware ensemble architecture that integrates statistical, sequential, and spatial representations of raw byte fragments. Extensive experimental analysis demonstrates that Triumvir consistently outperforms state-of-the-art methods with gains of up to +4.5pp in binary and +6.4pp in multiclass classification. Ablation studies confirm that combining modalities is critical, yielding improvements of up to +5pp over partial configurations.

The IETF standard Manufacturer Usage Description (MUD) enables manufacturers to equip IoT devices with certified URLs that provide traffic profiles for those devices, helping administrators enforce network access control. However, MUD assumes devices operate on full IP stacks and therefore does not account for constrained IoT devices running Thread–the dominant low-power mesh networking standard–which lacks complete TCP/IP functionality. While prior work proposes extensions to support MUD in Thread environments, these approaches are limited to simple topologies with a single border router and do not scale to realistic deployments with multiple, heterogeneous border routers. We introduce MeshGuard, a framework enabling MUD-based access control in complex Thread networks, with any number of border routers. MeshGuard extends the Mesh Link Establishment (MLE) protocol to deliver MUD information from constrained devices to border routers regardless of network topology. Moreover, MeshGuard leverages Software-Defined Networking (SDN) to synchronize access control lists across all routers. Experiments on our proof-of-concept with real devices (nRF5340, nRF52833, Raspberry-Pi 3) demonstrate enhanced security, minimal overhead, and linear scalability compared to state-of-the-art approaches.

Traditional traffic analysis is being fundamentally challenged by the rapid adoption of encryption, tunnelling, and privacy-preserving protocols, which increasingly obscure packet payloads and limit the usefulness of Deep Packet Inspection (DPI). Although machine learning has advanced encrypted traffic analysis, existing approaches often remain tied to protocol-specific header features, depend on large labelled datasets, and degrade when deployed across heterogeneous network environments. We present GETA, a protocol-agnostic framework for encrypted traffic analysis that models network flows as multivariate time series using only traffic metadata, thereby avoiding reliance on packet payloads or header semantics. GETA combines meta-learning, embedding refinement, and self-attention to support few-shot adaptation to previously unseen domains with minimal labelled data. Across nine public datasets spanning application identification, VPN traffic classification, IoT device fingerprinting, and attack detection, GETA consistently outperforms state-of-the-art baselines. These results show that GETA offers a practical and generalisable foundation for robust traffic analysis in modern encrypted networks.