CS Paper Radar

Daily listing

2026-11-17 · cs.CR

27 papers first submitted this day (UTC) in cs.CR — Cryptography and Security. arXiv announces a day's submissions the next evening, so the newest day fills up over ~24 h.

Score = weighted mean ofCitations 20%Upvotes 10%Code stars 10%Venue 15%Code/data 10%Novelty 15%Momentum 15%Substance 5%how it works

AI & Machine Learning 3

Distortion-Free Semantic Watermarks for Black-Box Generative Models

cs.LG cs.CRcs.CV ·2026-11-17 ·Aviv Yossef, Or Patashnik, Or Zamir
code/data
Abstract · Project page: https://avivyossef.github.io/distortion-free-semantic-watermark/ Code: https://github.com/avivyossef/distortion-free-semantic-…

We present a provably distortion-free semantic watermarking algorithm that requires only black-box sampling access to a generative model and black-box access to a semantic embedding. A watermark is called distortion-free if, averaged over its secret detection key, it leaves the generator's output distribution unchanged; this provides a formal guarantee that watermarking does not reduce output quality. Conventional such watermarks are often brittle to transformations. This is especially problematic for images, where benign operations such as compression, cropping, or resizing change the entire pixel representation. Semantic watermarks therefore require detection to be robust to content-preserving transformations. Previous methods combining these properties rely heavily on specific white-box properties of the underlying model: either to autoregressive text generation or to diffusion-specific access, such as control of the initial noise or diffusion inversion. In contrast, our construction is fully generic: It only interacts with the generative model and the semantic embedding, a function that determines the notion of "preserving content" under which detection should be robust, via sample and query access. We then instantiate this framework for image generation. We characterize the two competing properties required of the semantic embedding: it should remain stable under transformations that preserve an image's content, but should distinguish independent generations from the same prompt. We train an image encoder directly for this tradeoff. The resulting encoder improves separation while maintaining stability over off-the-shelf alternatives and enables strong aggregate detection while remaining robust to various transformations. We demonstrate the effectiveness of our watermark scheme on both open-source and closed models, which is possible only due to restriction to black-box access.

0.37

Who Pays for Byzantine Robustness? Validity, Data-Share Dictatorships, and Fairness When Only a Few Hospitals Collaborate

cs.LG cs.CR ·2026-11-17 ·Adnan Malkani
benchmarkcode/data
Abstract · 12 pages, 7 figures, 6 tables. Code: https://github.com/adnanmalkani/byzantine-fl-bench

Clinical federated learning (FL) collaborations typically connect only three to six hospitals, yet Byzantine-robust aggregation rules are designed and benchmarked on synthetic partitions across dozens of equal-sized clients. We conduct an empirical audit of eight standard aggregation rules: FedAvg, coordinate-wise median, trimmed mean, Krum, Multi-Krum, Bulyan, geometric median (GeoMed), and Centered Clipping (CC). We evaluate both unweighted and sample-size-weighted variants across three cross-silo medical datasets from the FLamby benchmark: Fed-Heart-Disease ($n = 4$ hospitals), Fed-TCGA-BRCA ($n = 6$ regions), and Fed-ISIC2019 ($n = 6$ centres). We execute 54,214 training runs across enumerated attacker coalitions ($f \in \{1, 2\}$) paired with exact-twin leave-out controls that decompose performance drops into honest data loss versus adversarial attack damage. We report three findings. First, client-count validity checks fail to capture breakdown in skewed medical consortia: while $n = 6, f = 1$ satisfies every count condition except Bulyan's, assigning sample-size weights turns median and GeoMed into a clean-model dictatorship whenever a single hospital holds more than half the training data, and causes weighted median, GeoMed, and CC to collapse when that dominant centre attacks. Second, exact-twin leave-out controls reveal that when the largest hospital attacks, the apparent failure of count-based robust rules is predominantly caused by the loss of honest training data (7.5 to 10.4 points on ISIC) rather than adversarial perturbation (2.3 to 3.8 points on ISIC). Conversely, weight-dependent rules suffer large true attack damage (16.1 to 25.1 points on ISIC). Third, unweighted robust rules consistently discount whichever institution generates the most atypical parameter updates, penalizing the smallest hospital on Heart but penalizing larger institutions on TCGA and ISIC.

0.36

ConfGuard: a physics-grounded reference monitor for predictive-maintenance large-language-model agents

cs.AI cs.CRcs.LG ·2026-11-17 ·Arunkumar V, Gangadharan G. R., S. Senthilkumar
Abstract · 24 pages, 5 figures

Large-language-model agents now close the industrial predictive-maintenance loop, reading sensor windows, retrieving procedures and dispatching work orders. Two attack channels meet there: injected text in a retrieved document can hijack the decision, and a bounded sensor perturbation can corrupt the remaining-useful-life estimate behind it. Recent agent defenses police data provenance and tool privilege but cannot tell whether a prognostic quantity is physically plausible. ConfGuard is a reference monitor with a physics-grounded policy. Its conformal monitor tests each prediction, anchored to the asset's admitted history, against degradation physics: capped remaining life never rises and falls at most one cycle per cycle. Calibrating its threshold over whole asset trajectories lets the operator set the expected share of a clean asset's windows the running monitor escalates. A physics-policy validator checks the agent's maintenance-urgency decision against the decision a deterministic policy derives from the same interval. On all 10,196 C-MAPSS test windows the monitor cuts a covert attacker's reach on healthy assets from 0.529 to 0.046 and closes suppression of failing assets (0.151 to 0), escalating 0.027 of clean windows against a 0.05 target, while the fixed physical range it replaces leaves both attacks open. Behind a Qwen3-14B planner the composed monitors hold urgency-class injection to 0.00 and the universal sensor attack to 0.12, against 0.29 and 0.54 for a defended prompt. An adaptive attacker defeats a Mahalanobis detector on 0.504 of windows and adversarial training on 0.485. We report an attack class the projection misses.

0.31

Language, Vision & Media 3

Propose with the Gradient, Accept with the Loss: One Hedged Attack for Pixel-Sparse Robustness

cs.CV cs.CRcs.LG ·2026-11-17 ·Hon Ming Yan, Chun Pong Lau
Abstract · 22 pages, 13 tables, 4 figures

A pixel-sparse adversary may change a few pixels of an image to any value. Robustness to such attacks is currently measured with an ensemble, sparse-AutoAttack (sAA), because its members fail in different places: sparse projected gradient descent (sPGD) trusts a first-order model that a full pixel change breaks, and sparse random search (Sparse-RS) ignores the gradient even where it is informative. We show that both attacks are local searches, each a proposal distribution plus an acceptance rule, and ask what such a search can learn from the gradient. A Taylor bound says what to measure, and measurements at the states the search visits give five rules. The sign of the gradient is its most reliable use at a single pixel and sets a new pixel's value. Its ranking of candidate pixels should be followed loosely, and its ranking of the pixels to remove, which is often more reliable, can be followed more closely. A pixel should not be proposed twice before a move is accepted, and acceptance should be decided by the loss alone. These rules define Gradient-guided Random Search (GRS), which runs once and costs one forward and one backward pass per step. On 13 CIFAR-10 configurations and 5 ImageNet models, GRS matches or beats the best single member of sAA and is within one point of sAA itself or stronger on 17 of 18 settings, and it is significantly stronger than sAA on 8. It uses 3 instead of up to 7 forward-pass units, and on CIFAR-10 it runs 3.7 to 5.0 times faster than the members of sAA run in full.

0.29

Fast and Strong Image Protection against Diffusion Personalization

cs.CV cs.CR ·2026-11-17 ·Jiwon Park, Jihye Kang, Namhyuk Ahn
Abstract

Personalized text-to-image diffusion models can reproduce a person from a handful of photographs. Image protection perturbs each image before publication so that fine-tuning on it fails, but it optimizes every perturbation from scratch, so its cost grows with the number of images. We present the protective diffusion model (PDM), which is trained once and samples the perturbation of a new image over many denoising steps. PDM distills an iterative teacher and is then post-trained with the target residual projection (TRP) loss as a reward, a stronger objective than existing ones. PDM outperforms nine baselines on faces and artworks and nearly matches its teacher at a tenth of the cost.

0.21

GuardAgent: A Memory-Augmented Agentic Framework for Adaptive Detection of Adversarial Promotions in LLM-Based Recommendation

cs.IR cs.CR ·2026-11-17 ·Shiyi Yang, Xinshu Li, Chen Wang, Xiwei Xu, Liming Zhu +1 more
published
Abstract · Accepted by WSDM 2027

In traditional recommender systems (RSs), textual information typically serves as auxiliary signals for representation learning. In contrast, in LLM-based RSs, textual information becomes the primary medium for recommendation generation through semantic understanding and reasoning. This shift fundamentally changes the attack surface, making promotion-driven manipulations more feasible, effective, and harder to detect. However, effectively detecting such attacks remains largely unexplored. To fill this gap, we present the first systematic study on detecting promotion-driven attacks against LLM-based RSs, aiming to safeguard next-generation recommendation platforms. Specifically, we propose GuardAgent, a novel LLM-powered agentic detection framework that operates through a loop of Perception, Action, Reflection and Memory. In contrast to conventional LLM-based detectors that function in a static, one-off, and opaque fashion, GuardAgent continually learns and self-evolves to adapt to unseen threat patterns while generating explainable reasoning traces and outcomes. Extensive experiments on real-world datasets demonstrate that GuardAgent achieves state-of-the-art generalization performance, laying the foundation for robust and trustworthy recommendation in practical settings.

0.20

Security & Privacy 19

From Attack Tree to Model Checker: Binding CAPEC Patterns to Verify Robotic System Security

cs.CR cs.LO ·2026-11-17 ·Juel Hussain, Louise A. Dennis, Clare Dixon, Marie Farrell
published
Abstract · In Proceedings FMAS 2026, arXiv:2611.21375

Security-minded verification of autonomous and robotic systems presupposes an attacker, yet the path from a threat catalogue to a checkable attacker model is still traversed by hand. This research preview presents the theory and construction of a pipeline that mechanises it. Each leaf of an attack tree is mapped to a CAPEC attack pattern and written in first-order logic as a rule: the attacker's prerequisites imply the attack's effect. Casting the pattern as such a rule is what makes it operational - its variables can be bound to a specific system's parts, and its prerequisites can then be tested against that system under a chosen configuration of security controls. Binding is our means of verifying whether a system is vulnerable to a given attack: from a bound pattern we derive an attacker model, turn its effect into a temporal-logic property, and check that property against the system with and without a countermeasure, reporting the control that defeats the attack. We describe the signature, schemas, profiles for approach, and explore how CAPEC can support reasoning about robotics security.

0.67

Is 4-Bit Quantization More Than Noise to a Chat Model's Safety Behaviour? Attributing Behavioural Shifts with Matched Perturbations

cs.CR cs.CL ·2026-11-17 ·Ege Kaya
code/data
Abstract · 16 pages, 4 figures. Code and labels: https://github.com/egekaya1/qrefusal

Chat models on phones and laptops are usually 4-bit copies of the model that was evaluated, and audits of such copies disagree about what quantization does to refusal. A 4-bit build is one of many weight perturbations of the same size, so a change in its behaviour could come from its specific rounding error or from any perturbation that large. In a pre-registered study of three 3-4B chat models, with paired tests, blind human labels and an 8-bit null, we find two 4-bit shifts of opposite sign: with a compliant reply prefilled, Qwen2.5-3B continues into harmful content more often, and on held-out jailbreaks Llama-3.2-3B complies 10.3 points less often. Matched perturbations separate the two. Shuffled and Gaussian versions of the error average close to zero, while scaling the real error, from its mirror image to twice its size, changes behaviour near-linearly. At half the real error, Llama's likelihood is almost unchanged, yet it already complies 5.8 points less often, and llama.cpp's Q4_K_M, at the same bits per weight, moves neither model. Refusal under 4-bit quantization depends on the specific rounding error, not on bits per weight or likelihood, so each build has to be compared with its original.

0.63

A Little Advice Suffices: Diffie-Hellman Is as Hard as Discrete Logarithm

cs.CR math.NT ·2026-11-17 ·Yuwen Zhou
code/data
Abstract · 15 pages. Developed with Claude Fable 5.2 and Claude Opus 5.5 (Anthropic); see "Use of AI". Companion paper on the hyperelliptic curve metho…

Is computing a Diffie-Hellman key as hard as computing a discrete logarithm? Maurer showed that it is, given a few bits of advice, provided that short intervals contain enough integers with only small prime factors; that proviso has stood for thirty years. We remove it. For every prime $p$ there is an advice string of $O(\log p)$ bits, depending only on $p$, with which discrete logarithms in every group of order $p$ are computed in polynomial time from a Diffie-Hellman oracle. The reduction is the classical one; what was missing was a proof that short intervals contain polylogarithmically smooth integers. We show that a zero-free half-plane for the Riemann zeta function, which is now known, gives the expected number of $(\log x)^K$-smooth integers in every interval $[x,x+x^θ]$ with $θ>17/30$. With smoothness this small the saddle point of the counting integral lies just to the left of the line $\mathrm{Re}\,s=1$, and a zero-free strip of any fixed width absorbs the shift once $K$ is large. Genus-2 Jacobians, whose orders fill such intervals, then turn one of these integers into the order of an auxiliary group. The same input removes the smoothness assumption from Maurer and Raub's impossibility of field-homomorphic one-way permutations and from the yes/no-query reduction of Koblitz, Menezes and Shparlinski. It also makes Wolf's characterisation unconditional: discrete logarithm and Diffie-Hellman are generically equivalent exactly when the group order has no large repeated prime factor.

0.62

Network-Exposed Configuration Attacks on LLM-Controlled Social Robots

cs.CR cs.AIcs.RO ·2026-11-17 ·Stanley Akingbola, Ugbede Ibrahim, Yan Lin Aung, Quang Dan Le, Hong Qing Yu
code/data
Abstract

Large language models (LLMs) in embodied robots create security dependencies beyond inference. We investigate a network-exposed configuration attack on the Pollen Robotics Reachy Mini Lite, allowing an unauthenticated attacker to replace personality settings and the cloud LLM's system prompt through a management interface, without adversarial conversational input. Eleven prompt-injection attempts across five categories were reported as resisted, whereas both configuration-layer attack paths produced observable behavioural changes on physical hardware. We demonstrate local and network-adjacent access, mapping the weakness to CWE-306 (Missing Authentication for a Critical Function) and CWE-1327 (Binding to an Unrestricted IP Address), with a reported CVSS v3.1 score of 6.5 (Medium). A configuration endpoint also accepts cross-origin requests without a CORS preflight (CWE-352; 6.5 Medium), while the persistent-memory sink stores content without enforcing its stated restrictions. A source review of 59 public repositories identifies missing authentication in 10 of 11 source-verifiable genuine applications, or 11 of 11 including a framework-default case. These findings indicate recurrence within the sampled repositories, not prevalence in commercial products. We characterise a deployment gap between developer-oriented network assumptions and consumer-facing embodied AI security: inference-layer refusal cannot prevent unauthorised changes to model instructions. Potential consequences include surveillance through onboard cameras and microphones and social engineering of nearby users. Proof-of-concept code, evidence and the survey are available at https://bit.ly/reachypi.

0.61

Can Verifiable PIR Match the Online Efficiency of SimplePIR?

cs.CR ·2026-11-17 ·Xinxuan Zhang, Ruida Wang, Zirui Wang, Yuming Guo, Jikang Bai +3 more
top venue
Abstract · Accepted for publication at the 2027 IEEE Symposium on Security and Privacy (IEEE S&P 2027)

Verifiable Private Information Retrieval (vPIR) ensures that a malicious server cannot return tampered database records, but it traditionally incurs significant overhead compared to classical PIR.Recent advancements, notably VeriSimplePIR, approach the online efficiency of SimplePIR by introducing a proof preprocessing phase and requiring only lightweight client-side online verification. However, this scheme relies on an additional Short Integer Solution (SIS) assumption and SIS-based commitments, which only support relaxed range proofs. Consequently, these schemes are forced to adopt significantly weaker parameter choices, resulting in additional online overhead. In this paper, we propose a novel vPIR protocol that employs a newly designed error-correcting code-based commitment scheme as the database digest. This approach (i)supports efficient proofs of linear relations, (ii) achieves precise range proofs and (iii) eliminates the reliance on the SIS assumption. Thus, our scheme successfully adopts parameter choices identical to those of SimplePIR. As a result, our vPIR protocol is the first to achieve exactly the same online communication and server computation costs as SimplePIR, while keeping online client computation nearly identical. Additionally, compared to VeriSimplePIR,our protocol reduces online communication and client-side computation by approximately 20%, and decreases online server-side computation by over 35%.

0.52

How Well Is My Artwork Protected? Predicting and Tailoring Protection against Style Mimicry by Personalization

cs.CR ·2026-11-17 ·Jihye Kang, Jiwon Park, Sungwoo Hwang, Namhyuk Ahn
unusual
Abstract

Protection tools add an imperceptible perturbation to each artwork so that a model personalized on it fails to imitate its style, and they apply one recipe, a protection method at a fixed strength, to every artwork. We find that the same recipe protects artworks very differently, well beyond the noise of personalization seeds, and that cheap signals predict this protection before any personalization. Because protection rises and falls together across recipes, our method, Patina, predicts the protection of every recipe from one short trial protection and gives each artwork the weakest recipe whose predicted protection reaches a chosen level. At the same level of visibility, Patina lets more artworks pass the protection threshold than any single recipe and warns the artist about artworks for which no recipe is predicted to pass it.

0.42

From Token Changes to Task Failures: Bit Flips in a Shared KV-Cache Block

cs.CR cs.AR ·2026-11-17 ·Yuji Yamamoto, Satoshi Matsuura
Abstract · 19 pages, 4 figures. Extended version of arXiv:2604.17249 (SECRYPT 2026, 23rd International Conference on Security and Cryptography). Submit…

In LLM serving systems with prefix caching, each KV block of a shared system prompt is held in GPU memory as a single physical copy, which every request with that prompt reads. Our earlier work measured change and showed that one bit flip in such a block changes many later responses without making them visibly degenerate. This paper extends that work by measuring how often the changed responses are wrong. Using software fault injection into vLLM serving Qwen3-8B, we flip the most significant mantissa bit or the sign bit of a value in such a block. We score responses on four tasks: code generation, short-answer questions, multiple-choice reasoning, and refusal under a policy. The fraction of changes that turned a correct response wrong depended strongly on the task: 0.023 on MBPP+ and 0.59 on held-out TriviaQA at the mantissa bit, and 0.019 and 0.60 at the sign bit. Harm fell on requests whose sampled answers are sometimes wrong, so the damage of a flip depends on how common such requests are. Harmed responses looked like ordinary errors, so a flip should be detected by checking cached blocks rather than by monitoring outputs.

0.40

Guarded Swarms in the Loop: A ROS 2/PX4 Digital Twin for Hybrid Digital-Analog Coordination of UAV/UGV Swarms

cs.CR cs.MA ·2026-11-17 ·Uwe M. Borghoff, Paolo Bottoni, Luca D'Aurizio, Remo Pareschi
Abstract · 10 pages, 1 figure, 7 tables. Submitted to ICMCIS 2027

Swarms of unmanned aerial and ground vehicles (UAV/UGV) operating in contested environments cannot rely on software alone for physical safety, since the software stack and its networks are part of the attack surface. The Guarded Swarms architecture addresses this by pairing a digital coordination layer, built on Topic-Based Communication Space Petri Nets (TB-CSPN), with an independent analog safety layer that can veto unsafe actuation commands. So far, the architecture has been specified and argued for, but not executed. We present the first executable instantiation of Guarded Swarms as a digital twin built on ROS 2, the PX4 autopilot in software-in-the-loop mode, and Gazebo, with two UAVs and one UGV in a coastal counter-UAS scenario. The TB-CSPN transition chain, from multi-modal sensing through human authorization to vehicle tasking, is mapped one-to-one onto ROS 2 nodes, and the analog layer is modeled as a per-vehicle Virtual Safety Board that gates every command. Eight groups of experiments show that the three negative guarantees of the architecture (no mission from conflicting, temporally incoherent, or unauthorized evidence) hold at three distinct stages of the chain in all 60 counterfactual runs; that threat confirmation is bounded by the narrowest sensing modality; that the safety layer exercises its full modeled envelope independently of the coordination layer; and, under fault injection, that it prevents every collision caused by a forged mission or a compromised worker once its veto is enforced by an independent gate, whereas a veto left to the compromised worker prevents none. The twin also exposes implementation-level hazards that the architectural specification does not address, and it draws an explicit line between the properties that simulation can validate and those that only hardware can.

0.39

TetherMark: Instance-Level Watermark Authentication Against Cross-Instance Transplantation

cs.CR cs.CV ·2026-11-17 ·Jin Wei, Rundong Li, Li Ma, Yikai Wang, Ruihao Yang +9 more
hot direction
Abstract

Robust watermarking is commonly verified by asking whether a hidden payload remains decodable after compression, geometric distortion, or generative editing. Localized and semantic extensions may additionally check whether an expected region or category is still present. These tests establish evidence validity or content consistency, but not whether the evidence still belongs to the particular visual instance registered by its owner. We expose cross-instance transplantation: an attacker extracts or estimates valid watermark evidence from one registered image and transfers it to a different image of the same category. The message can still decode, the expected content can remain present, and the result can look natural, even though the evidence has been reassigned to the wrong instance. To enable instance-level watermark authentication, we introduce TetherMark, a post-hoc watermark that combines two complementary checks. The message check verifies an edit-robust payload, while the relation check uses secret patch-pair selection, a registered response, natural correspondence transport, and enrollment normalization to test consistency with the registered instance. The verifier accepts only when both checks pass. Across 240 test subjects, TetherMark attains 92.08% legitimate true acceptance with 9.17% and 16.25% attack success under WmForger and WMCopier, respectively, yielding the highest instance-binding balanced accuracy under both attacks.

0.38

Provable Mixing Properties of Parallel Mixnets Revisited

cs.CR ·2026-11-17 ·Pierfrancesco Ingo, Peter Druschel, Deepak Garg
Abstract · 16 pages, 3 figures, 2 pages of appendix

A mix-network (mixnet) is a routing network that breaks source-to-destination message trackability by indistinguishably mixing the routes of concurrently transmitted messages, thus providing anonymity to senders, receivers and sender-receiver pairs. Given their potential use in high-stakes scenarios like protecting whistleblowers, one would like to prove formally that sender anonymity holds in a mixnet, even against strong adversaries that can compromise subsets of mixnet servers. However, existing work on anonymity in mixnets is either limited to linear mixnets (which do not scale horizontally), relies on special provisions such as additional noise or chaff messages, or is empirical -- typically based on entropy, which does not provide strong guarantees if the adversary has auxiliary information. In this paper, we present a new analysis of parallel mixnets which scale horizontally, showing that these networks can provide strong indistinguishability for messages even without additional noise messages. Our analysis works by bounding the statistical variance of the adversary's posterior likelihood of where a message of interest is at a given point in time. Using concentration bounds on this distribution, we show that parallel mixnets satisfy $(\varepsilon, δ)$-differential indistinguishability for messages routed through the mixnet (similar to $(\varepsilon, δ)$-differential privacy) with very small failure probabilities $δ$ and small privacy parameters $\varepsilon$ under reasonable operating conditions. Our model also enables an examination of the quantitative trade-off between anonymity, latency and per-router bandwidth in parallel mixnets.

0.37

Exploited Before It Was Fixed? Vulnerability Timelines from Open, Machine-Readable Evidence

cs.CR ·2026-11-17 ·Cédric Bonhomme, Alexandre Dulaunoy
Abstract · 8 pages

-05-X-03), Known Exploited Vulnerabilities (KEV) catalogs, and exploitation sightings. We introduce an evidence model implemented in Vulnerability-Lookup that selects the earliest evidence for each lifecycle event, distinguishes vulnerability-specific records, preserves date precision, and ensures traceability to original sources. Among 3,758 vulnerability identifiers with exploitation evidence, 98% have disclosure dates, while only 15.7% have patch evidence. Vendor-side dates required by VTEM are entirely absent. Most exploitation dates originate from catalog listings, including 1,041 entries associated with bulk catalog imports. After excluding these entries and ambiguous same-day observations, 59 vulnerabilities exhibit patch evidence following exploitation evidence, with a median interval of 7.6 days, predominantly between catalog listings and distributor fixes. Evidence traceability also exposes systematic measurement errors: one collector incorrectly classified exploit publications as active exploitation, accounting for 40% of initial exploitation dates, while another catalog assigned vulnerability publication timestamps to 11% of its listings. Our findings demonstrate that exploitation-to-patch intervals are computationally measurable, but their reliability depends on the availability, precision, and provenance of published evidence. We identify the additional disclosure, exploitation observation, and patch availability timestamps needed to improve vulnerability lifecycle measurements.

0.30

GapLeak: When Privacy-Preserving LLM-based Web Agents Become a Privacy Risk

cs.CR ·2026-11-17 ·Chaklam Cheong, Tiancong Cao, Qi Li, Zelin Zhang, Jianbing Ni +2 more
published
Abstract · Accepted for publication in Proceedings on Privacy Enhancing Technologies (PoPETs), 2027, Issue 2 Abstract shortened for arXiv character lim…

Large Language Model (LLM)-based web agents rely on data minimization and least privilege to protect user privacy. We show that these principles are necessary but insufficient: they constrain authorized sources, but not higher-sensitivity sources that remain reachable. A website can request a field unavailable from authorized sources, creating an information gap. To complete the task, a goal-driven agent may escalate to reachable but unauthorized sources and disclose accumulated sensitive information. We formalize this vulnerability as GapLeak, where an adversary exploits a legitimate web form without prompt injection or other malicious instructions. We evaluate GapLeak on three web-agent frameworks and six LLMs from Anthropic, Google, and OpenAI across recruitment, e-commerce, travel, and healthcare. GapLeak induces an average unauthorized disclosure of approximately 37%. The information gap is the primary driver: eliminating it reduces disclosure by more than 50%. Existing instruction-level defenses based on least privilege partially mitigate the vulnerability but often reduce task success. We propose submission-time verification, which filters unauthorized content before form submission, eliminating over 96% of unauthorized disclosures while preserving task success.

0.30

Dressed in Authority: Practical Black-Box Jailbreaking of MLLMs via Visual Trust Bias

cs.CR ·2026-11-17 ·Weihang Liu, Qianqian Xu, Yangbangyan Jiang, Zitai Wang, Zhipeng Yu +1 more
Abstract · 10 pages, 9 figures

Multimodal large language models (MLLMs) are rapidly advancing, yet their safety remains a critical concern, especially under jailbreak attempts. Existing MLLM jailbreaks either inject fragile adversarial perturbations into irrelevant images or synthesize toxic visual cues to achieve text-driven attack intents. In contrast, a more practical MLLM jailbreak emphasizes enhancing attack efficacy directly on naturally occurring harmful image-text pairs. We address this problem by proposing Visual Trust Bias Attack (VTBA), a new jailbreak framework that exploits a previously overlooked attack surface: MLLMs exhibit an intrinsic trust bias toward authoritative visual features, which enables transferable and stealthy jailbreaks. VTBA constructs a bank of authoritative templates, matches each input pair to the most relevant one, and integrates the harmful intent into the selected layout, thereby leveraging credible visual cues to override safety refusals while preserving naturalness. We further construct a dataset of 2,500 naturally occurring image-text pairs to evaluate jailbreaks in this practical multimodal setting. Extensive experiments show that our method consistently achieves high attack success rates across strong MLLMs, revealing a severe and previously underexplored weakness in current multimodal safety alignment. Warning: This paper may contain sensitive content.

0.29

Adaptive LLM-Driven Honeypots vs. LLM-Driven Attackers: Does Narrative Coherence Improve Deception?

cs.CR cs.AI ·2026-11-17 ·Pedro Beltrán-López, Raul Gonzalez Jimenez, Dario Maddaloni, Manuel Gil Pérez, Pantaleone Nespoli
Abstract

Large Language Models (LLMs) enable honeypots to generate responses and deceptive artefacts on demand, yet existing approaches evaluate generated content one reply at a time and leave open whether artefacts that jointly describe a coherent event retain an attacker better than unrelated ones. This question becomes pressing as autonomous LLM-driven offensive agents emerge as realistic adversaries. This paper proposes narrative deception, a reactive strategy in which the defender maps each observed attacker command to a MITRE ATT&CK technique, selects a compatible scenario from a curated catalogue based on MITRE Engage, and deploys mutually referential artefacts (files, processes, services, and logs) on an emulated Ubuntu host while the intrusion is under way. An LLM-powered defender and an autonomous LLM attacker interact through a real SSH session, and the attacker's termination policy distinguishes abandonment due to suspected deception from abandonment due to loss of interest. A campaign of 1,150 sessions compares four conditions of increasing sophistication (Static, Same-TTP, Multi-TTP, and Chain) across four open-weight attacker models, four defender models, and three attacker skill levels. Engagement increases significantly at every step. With the same artefact budget per reaction, coherent scenarios raise the median session length of advanced attackers from 40 to 165 commands and reduce their abandonment from 86% to 52%. The effect grows with attacker skill, holds across all sixteen model pairs, and comes at comparable inference cost. Nearly all sessions ended through loss of interest rather than detection, which suggests that deception should sustain meaningful findings, not only emulator fidelity.

0.26

Flexibility Has a Price: Quantifying Modality Policies in Maliciously Secure Multi-Biometric Authentication

cs.CR ·2026-11-17 ·Fuyi Wang, Wang Zhang, Neo Xu, Chao Chen, Kok-Leong Ong +1 more
Abstract

Multi-biometric authentication improves robustness and accessibility by combining information from multiple biometric modalities. Existing research has primarily focused on secure multimodal authentication, feature fusion, and recognition performance, while paying comparatively little attention to the security implications of modality-flexible authentication policies. At the same time, most systems provide limited cryptographic protection for biometric data during authentication. This paper addresses both challenges. We first study modality flexibility as a security decision rather than merely a usability feature. Each accepted modality subset creates an authentication entry point, and the resulting risk depends on the joint behavior of all accepted subsets and the attempt budget. We introduce a framework for quantifying this security cost and develop a policy-selection procedure that identifies policies satisfying a target risk level while supporting users who can present only a subset of modalities. We then present SUPERMBA, a two-server privacy-preserving authentication system that enforces the selected policy without revealing biometric templates or the modality subset used in an authentication attempt. The protocol provides verifiable policy enforcement even when one server is malicious and executes the same computation irrespective of the chosen policy or modality subset. Experiments on the three-modality MULB+ dataset show that allowing all modality subsets increases impostor risk from 1.21% for the weakest entry to 4.42% for the union of entries. On VidTIMIT, the increase is substantially smaller and is not statistically resolved with the available identities. SUPERMBA completes an authentication attempt in only 6.42 ms on MULB+ and 6.03 ms on VidTIMIT.

0.26

Stopping Is Not the Limit: Understanding and Eliciting Agent Capability in Long-Horizon Vulnerability Exploitation

cs.CR cs.AIcs.SE ·2026-11-17 ·Deokhyung Kang, Minjae Gwon, Sangdon Park
Abstract · Preprint

Vulnerability exploitation tests whether language model agents can develop working exploits from known software flaws. Sequential test-time scaling can improve exploitation performance, but larger budgets do not ensure sustained exploration: agents often stop with budget remaining. Yet what happens when execution continues beyond stopping remains underexplored. We investigate this gap on ExploitBench using DeepSeek-V4-Flash-0731 and MiMo-V2.6-Pro. We find additional verified gains after the first stop response, but gains diminish over longer execution. We identify two patterns associated with less productive continuation: an increasing proportion of stop responses within the same context and earlier stopping associated with summaries retaining impossibility judgments about unachieved capabilities. To test whether context management can elicit further capability, we propose SCoRe, which compacts context at stop bursts and marks these judgments as unproven. SCoRe improves verified capability over summary-based continuation and other evaluated context-management baselines. With SCoRe, MiMo reaches a capability coverage score of 67.0% at 2,000 turns, exceeding both 32.4% under the original stopping policy and the 58.6% computed for GPT-5.5 from ExploitBench's 300-turn reference results on the same tasks. These findings show that stopping need not mark an agent's exploitation limit, and that continued execution with context management can elicit substantial further capability.

0.22

SEAM: Computer-Use Agents Know the Harm Between Screens, and Commit It Anyway

cs.CR cs.AI ·2026-11-17 ·Yuxuan Zhou, Chen Xiong, Haitong Liu, Kuofeng Gao, Yang Bai +3 more
benchmark
Abstract · 30 pages, 14 figures, 29 tables

Safety benchmarks for computer-use agents (CUAs) place the harm either in the intent of the task or on a single screen: a malicious instruction, an injected command, a phishing page. An agent that checks the task and each screen in isolation can pass them, so they do not test whether it combines facts seen on different screens. Yet in delegated work, harm often arises exactly this way: a fact about a person on one screen, a fact about an item on another, and a legitimate task that binds them. We introduce SEAM (Safety Execution Across Multiple screens), a benchmark of such cross-screen compositional harm: 600 multi-step, cross-application OSWorld tasks in which the instruction and every single observation are benign and only their combination is harmful. Minimal-pair variants that differ in one place, scored from the server log rather than by an LLM judge, locate where an agent fails. General-purpose open-weight and flagship agents carry out the harm in up to 92% of runs, yet decline when the user states it in one sentence. Agents open the decisive pages, so exposure is not the primary bottleneck. Asked about those pages, they flag the harm, and with thinking they often write the conflict into their reasoning and confirm anyway. Safety recognition thus does not reliably translate into safe execution, and the translation depends on the information channel: the same warning stops agents far more often in the user's instruction than on screen. Proprietary models are markedly more cautious. We release SEAM with its generator and quality gates at https://yuxuan2003.github.io/SEAM/ .

0.22

CaFe: Causally Faithful Conditional Traffic Scenario Generation

cs.CR cs.CVcs.RO ·2026-11-17 ·Steven Yu Tung, Shubham Vaijanath Phoolari, Christoph Lauer, Abhinav Valada, Steven Peters
Abstract

Language-conditioned traffic simulation enables scalable test scenarios steered by natural language. The goal is to control selected agents while surrounding traffic reacts to their observed behaviour. For example, a vehicle instructed to cut in late should cause the following vehicle to brake only when the manoeuvre becomes observable. Yet simulators can leak private instructions across agents, causing surrounding vehicles to react before the instructed behaviour becomes observable. Such premature reactions defuse desired events and alter the ensuing closed-loop scenario. Existing benchmarks measure what agents do, not what they could have known, leaving this failure unmeasured. We formalise it as intent leakage and address it through evaluation, data and modelling. Our paired intervention test holds an instructed agent's executed trajectory fixed while varying its private instruction, isolating leakage-induced changes in surrounding agents. We also construct WOMD-PACT, an automatically generated, verifiable dataset of 531,099 WOMD scenes with timed behaviour tags with reaction episodes identified through a counterfactual ghost test and grounded private instructions. Finally, we introduce CaFe, a dual-stream diffusion-forcing simulator fine-tuned in closed loop that restricts inter-agent interactions to public beliefs, preserving legitimate anticipation without leaking private intent. Our evaluation reveals that existing conditioning designs displace 23-83% of nearby uninstructed agents through private-intent leakage, even when the instructed agent's executed motion is held fixed. CaFe eliminates measurable leakage while maintaining comparable WOSAC realism (0.763), demonstrating that causal faithfulness can be enforced without sacrificing measured simulation realism.

0.21

Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors

cs.CR cs.AIcs.CL ·2026-11-17 ·Yujun Wang, Yilun Liu, Jinhe Bi, Krishna Kanth Nakka, Xuebing Zhou +2 more
benchmark
Abstract

Tool-using language-model agents must act on information from untrusted emails, webpages, and documents. This creates a central security challenge: malicious instructions and legitimate task directives can arrive through the same interaction channel. Existing defenses are largely hand-designed and static, making further improvements dependent on repeated manual redesign. We introduce Evolution Under Fire, a staged framework that couples safety-harness evolution with adaptive stressor evolution. Utility-constrained, Pareto-scheduled tree search evolves complete executable safety harnesses from execution evidence while keeping the task model, benchmark, tools, and evaluator fixed. Adaptive stressor evolution then discovers utility-preserving attacks that guide a subsequent stage of harness evolution. On AgentDyn with GPT-4o-mini, an evolved harness reduces attack success from 50.4% to 3.0%, improves clean and attacked utility over the initial harness, and offers a stronger security-utility trade-off than reproduced defenses. On Agent-SafetyBench with Qwen2.5-7B-Instruct, harness evolution reduces attack success from 82.6% to 62.0% while retaining helpfulness and runtime validity. Evolved harnesses also transfer without further evolution to Qwen3.5-35B-A3B, reducing attack success from 32.1% to 0.9% while improving attacked-task utility.

0.11

Systems & Networking 1

Dissecting the GPU Load/Store Unit and Exposing Its Security Risks

cs.AR cs.CR ·2026-11-17 ·Wenbo Han, Pierre Talbot
Abstract

GPUs are the main platform for large-scale machine learning and power the largest supercomputers. Their load/store units (LSUs) often determine kernel performance and have been exploited in prior attacks. Yet what is known about the LSU is only a few isolated mechanisms and parameters, and even a single one has been enough for an attack. In particular, it is still unknown what a load costs, where loads contend, and even whether a warp can starve due to contention. Until each is measured, neither its security risk nor a defense is known. We reverse-engineer how the LSU serves load instructions on recent NVIDIA GPUs with microbenchmarks and analyze the security implications. We give a cycle-accurate latency formula for a global L1 hit or shared memory access from the addresses it touches, whose address-dependent terms are bank conflicts and the L1 tag lookup; we show that loads contend at both, and that when the LSU pipeline is full a warp can starve a warp of a later-launched block. We then show that the LSU's attack surface is far broader than prior attacks revealed: through channels never exploited before, it can leak information, communicate covertly, and degrade colocated workloads. Through either global bank conflicts or tag lookups alone, we recover a full AES-128 key under the threat model of prior GPU AES timing attacks. Under NVIDIA's Multi-Process Service (MPS), the bank access and the tag lookup each open a covert channel across processes, whose rate rises with each GPU generation. Also under MPS, this starvation lets one warp slow a colocated cuBLAS matrix multiplication by up to 750x on the A100 and 563x on the H100.

0.06

Software & Programming 1

Missing Security-Fix Backporting for GitHub Actions Workflows

cs.SE cs.CR ·2026-11-17 ·Chengyan Ma, Jieke Shi, Ruidong Han, Weifeng Sun, Yuqing Niu +3 more
top venue
Abstract · Accepted by the 49th IEEE/ACM International Conference on Software Engineering (ICSE 2027)

GitHub natively supports workflow automation through GitHub Actions, now the dominant CI/CD tool on the platform. Yet keeping these workflows secure remains an ongoing burden: maintainers often patch a vulnerability on the default branch but leave the same flaw on the release and maintenance branches that downstream consumers continue to build from. Carrying a known fix onto these divergent branches, called backporting, is rarely done, and this paper presents the first large-scale empirical study quantifying the resulting gap. From a public corpus of GitHub Actions workflow histories, we mine 1,903,191 workflow-modifying commits, scan each before and after with the static scanner zizmor to identify 69,865 security fixes across 24,238 projects, and audit every release branch to check whether each fix was carried over. The gap is pervasive: 224,392 of 474,490 audited release branches remain vulnerable, and deliberate backporting is rare and slow, with only 19,476 of 78,907 already-fixed branches confirmed as plausible backports at a median lag of 243 days. We thus design WorkflowBP, the first technique for automated security-fix backporting of GitHub Actions workflows. WorkflowBP takes a neuro-symbolic approach: it compiles a default-branch fix into a domain-specific representation called the Workflow Semantic Patch (WSP) and applies it to each divergent branch by matching jobs and steps based on their content rather than their position, while an LLM synthesizes the WSP for the most divergent cases the deterministic engine cannot resolve. On 12,000 real unpatched workflow files from our study, WorkflowBP produces a scanner-checked backport for 67.9%. We further submit 176 issues with WorkflowBP-generated backports to real-world projects, 44 of which have been confirmed or fixed by developers.

0.37