SEARCH · Searcharxiv

Search Searcharxiv

Search indexed arXiv papers on artificial intelligence, large language models, computer vision and robotics. Read source abstracts and follow links to arXiv.

Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.

At least 253 records · Page 14Linked to original sources

arXiv2026

Safety Training Modulates Harmful Misalignment Under On-Policy RL, But Direction Depends on Environment Design

Specification gaming under Reinforcement Learning (RL) is known to cause LLMs to develop sycophantic, manipulative, or deceptive behavior, yet the conditions under which this occurs remain unclear. We train 11 instruction-tuned LLMs (0.5B-14B) with on-policy RL across 3 environments and find that model size acts as a safety buffer in some environments but enables greater harmful exploitation in others. Controlled ablations trace this reversal to environment-specific features such as role framing and implicit gameability cues. We further show that most safety benchmarks do not predict RL-induced misalignment, except in the case of Sycophancy scores when the exploit relies on inferring the user's preference. Finally, we find that on-policy RL preserves a safety buffer inherent in the model's own generation distribution, one that is bypassed during off-policy settings.

cs.LG↗

arXiv2026

Sealing the Audit-Runtime Gap for LLM Skills

Large language model (LLM) ecosystems such as Claude Code and ChatGPT increasingly rely on skills: packages of natural-language instructions and executable tools. Once in the LLM's context, skill content cannot be reliably separated from trusted instructions, and a skill's executable side can invoke privileged actions, exposing the skill supply chain to injection, tampering, and rug-pull attacks. Existing defenses are stage-bound: centralized signing, audit reports unbound from the runtime artifact, or policy engines that cannot attest to what was approved. We present SIGIL, the first framework that seals the audit-runtime gap for LLM skills. SIGIL delivers verifiable hosting through a tamper-evident, decentralized on-chain registry from which LLMs fetch skills directly. The registry admits four publication types, Transparent, Licensed, Sealed, and Committed, spanning plaintext public distribution, monetized access, custodial use, and off-chain workflows; before admission, every skill is vetted by a Decentralized Autonomous Organization (DAO) audit committee that supports pluggable auditing methods under a stake-and-slash economic model. At load time, SIGIL delivers verified loading through a skill verification protocol executed by a Skill Verification Loader (SVL) embedded as the mandatory loading path: the SVL retrieves and decrypts the skill as its type requires, verifies its integrity against the on-chain record, and enforces its permission manifest before context injection. We evaluate SIGIL on a real-world deployment against 1,023 in-the-wild skills spanning six attack types. At load time, the SVL verifies each skill's integrity against its on-chain record and enforces its approved permission manifest, completing batched verification under 86 ms. Together, these results show that LLM skills can be cryptographically bound from publication through runtime at practical cost.

cs.CR↗

arXiv2026

Identifying AI Web Scrapers Using Canary Tokens

From pre-training to query-time augmentation, web-scraped data helps to improve the quality and contextual relevancy of content generated by large language models (LLMs). However, large-scale web scraping to feed LLMs can affect site stability and raise legal, privacy, or ethics concerns. If website owners wish to limit LLM-related web scraping on their site, due to these or other concerns, they may turn to scraper access control mechanisms like the Robots Exclusion Protocol. To be most effective, such mechanisms require site owners to first identify the scrapers that they wish to restrict (e.g., via User-Agent strings). Existing mechanisms to identify LLM-related scrapers rely on voluntary disclosure by companies, one-off experiments by researchers, or crowd-sourced reports -- methods that are neither reliable nor scalable. This paper proposes a novel technique for accurately and automatically inferring LLM-related scrapers. We host dynamic websites that serve unique canary tokens to each visiting scraper, then prompt LLMs for information about our sites. If an LLM consistently generates outputs containing tokens unique to a scraper, it provides evidence of exposure to that scraper. Via experiments across 22 production LLM systems, we demonstrate that our approach can reliably identify which scrapers feed which LLM, including several that are not publicly known or disclosed by the companies. Our approach provides a promising avenue for unprivileged third parties to infer which scrapers serve data to which LLMs, potentially enabling better control over unwanted scraping.

cs.CR↗

arXiv2026

Deep-Research Agents Can Be Poisoned via User-Generated Content

Deep-research agents are an alternative to conventional Web search. They use multi-agent pipelines to issue multiple Web searches related to user queries, retrieve relevant content from the answers, and generate detailed, evidence-based reports. We show that for many common search topics (including financial, medical, and product recommendations), agent-generated reports are consistently based on the same user-generated content (UGC) pages from platforms such as Reddit and Wikipedia. This retrieval overlap, combined with lax moderation and access controls for UGC, is an opportunity for subversion: an attacker who appends a short text to a single, frequently-retrieved page can cause agents to cite this text and promote attacker-chosen entities across many user questions. We evaluate this attack on representative deep-research systems across multiple topics, such as investment advice, antivirus software, restaurant recommendations, etc. Poisoning a single URL with as few as 13 words can be sufficient for the attacker's content to be retrieved in 57-76% of the agent executions for a given topic and cited in 38-51% of the generated reports. A more aggressive attack (poisoning an entire subreddit) achieves 30-53% citation rates even when the poison is only 0.5-4% of the retrieved content. The attacker does not need to know the phrasing of the user's question, nor the specific Web queries generated by the agent, nor the agent's internal retrieval and generation mechanisms. We then study defenses at different stages of the pipeline, including source-level filtering and output-based detection. Dropping UGC from retrieved content blocks the attack but degrades the quality of generated reports. We show that lightweight anomaly detection on inputs and outputs does not reliably identify poisoned content, nor the results of poisoning.

cs.CR↗

arXiv2026

Refusal Before Decoding: Detecting and Exploiting Refusal Signals in Intermediate LLM Activations

In this paper, we investigate whether refusal behavior can be predicted from LLM intermediate activations before decoding using linear probes trained on residual stream activations at each transformer block. We find that refusal is linearly decodable well before the final layer, indicating that safety-relevant behavior is represented in intermediate activations before output generation. To test whether this signal is actionable, we introduce Mechanistic AutoDAN, a probe-guided variant of AutoDAN that replaces full-model fitness evaluation with partial forward passes and probe-based scoring inside a genetic prompt search loop. Across the evaluated models, our method achieves attack success rates competitive with vanilla AutoDAN while reducing per-iteration search time by up to 72%, and probe-guided prompts match or exceed AutoDAN's cross-model transfer in several configurations. We further find that the usefulness of probe guidance increases with model scale. Our results suggest that refusal-relevant information is decodable from intermediate activations and can serve as an effective search signal in an AutoDAN-style discrete jailbreak optimization loop, especially for larger and more robust models.

cs.AI↗

arXiv2026

Robust and Efficient Guardrails with Latent Reasoning

Maintaining the safety of large language models (LLMs) is crucial as they are increasingly deployed in real-world applications. Existing safety guardrails typically rely on single-pass classification or, more recently, distilled reasoning. Reasoning-based guardrails significantly outperform classification-only baselines, but they incur substantial query latency and token overhead that make them impractical for highthroughput deployment. To address this challenge, we propose COLAGUARD, a guardrail model that transfers multi-step safety reasoning into a continuous latent space through a stage-wise training curriculum, enabling direct hidden-state propagation at inference. Evaluated on ten prompt- and response-moderation settings spanning eight safety benchmarks, COLAGUARD improves macro-F1 by 8.24 points over Llama Guard 3 and matches our explicit reasoning baseline, GuardReasoner, in macroF1 while delivering a 12.9X speedup and 22.4X reduction in token usage. Our results suggest that latent reasoning offers a practical alternative to explicit rationale generation for deployable guardrails, jointly improving safety robustness and inference efficiency rather than treating them as competing objectives.

cs.AI↗

arXiv2026

Longitudinal Adoption and Deprecation of the Privacy Sandbox Web APIs

While several web actors have been trying to reduce web tracking for years, it remains unclear how to achieve both desirable levels of utility and privacy. In 2019, Google launched the Privacy Sandbox initiative to balance that trade-off and find privacy alternatives to common use cases such as advertising. Yet, in late 2025, Google canceled the project and deprecated most of the newly introduced APIs. Despite its end, the Privacy Sandbox represents a unique opportunity to learn about how the ecosystem reacted to the proposed changes. In this paper, we present a longitudinal measurement and analysis study of the Privacy Sandbox APIs to characterize their adoption and deprecation over the past seven years by different web actors. Leveraging historical HTTP Archive crawls and public Chrome telemetry data, we offer the largest study of its kind into the prevalence of each Privacy Sandbox feature, during their entire respective lifetime (5+ years for some), on popular websites (CrUX top 100k), and as experienced by Chrome users during their browsing journey. Our results showcase an adoption that remained limited and uneven across the years; only few web actors implemented very specific APIs, and in disparate manners. We motivate our interpretation of these results by considering the incentives (interest, resources, timeline, etc.) and risks (potential trade-offs, privacy violations, and legal exposure, etc.) for these actors. Finally, our analysis also yields a few actionable recommendations for the next generation of web privacy-enhancing technologies.

cs.CR↗

arXiv2026

How Reliable Is the Multi-Input Heuristic for Bitcoin Address Clustering in Law Enforcement Contexts?

Address clustering is an important technique in blockchain forensics, widely employed by law enforcement to trace illicit crypto asset flows. The multi-input heuristic (MIH), which clusters addresses potentially associated with the same entity, is the most widely used. Yet, despite its broad adoption, the MIH has rarely been evaluated against reliable ground truth data. We implement a reusable evaluation framework covering nine established metrics and apply it to ground truth address-to-entity mappings obtained directly from European crypto asset service providers under legally mandated reporting obligations. When evaluation is restricted to reported addresses, the MIH appears strong at dataset level: we observe no mergers between reported services and recover same-service address pairs with recall 0.71. However, this result is driven by one large service and ignores unlabeled addresses absorbed into full clusters. Metrics that assess the full clusters show substantially lower precision and recall (0.36 and 0.44), meaning that services are often only partially recovered or embedded in larger clusters. Entity-level results further reveal near-complete failures for some services. When MIH-based clusters are used to support criminal suspicion, preliminary seizure of crypto assets to secure later forfeiture/ confiscation, or as evidence in trial proceedings, prosecutors and judges must account for the heuristic's metric-dependent and entity-dependent reliability.

cs.CR↗

arXiv2026

Semantic Overlays: Mitigating Prompt Injection with Annotations Beyond Tokens and Steering Vectors

Everything a language model sees is tokens. The serving stack knows what each span is -- user input, tool output, instructions -- but the model must keep track of that itself, and can lose track or be confused: text can be written to read like anything. Prompt injection is a natural exploit of this phenomenon. By scrambling the model's understanding of span identity, an attacker can induce unwanted and dangerous actions. Adding a non-textual channel to the model's input -- a way to communicate span identity beyond text -- mitigates this class of attack. We thus introduce a general steering technique called Semantic Overlays: small learned adapters applied at chosen prefill positions to a frozen model's residual stream. Laying an overlay over a span creates an out-of-band annotation channel that cannot be replicated by tokens. Unlike steering vectors, Semantic Overlays are trained, adaptable, and selectively applied. An overlay can encode complex semantics that reshape how the model perceives the marked span: asked to copy a code snippet under an overlay asserting a different programming language, the model rewrites the snippet in the asserted language. Overlays compose, allow transparent reading of underlying content, and can carry complex payloads -- including imperatives the model will follow. An overlay which marks a span as "non-executable" defends against the broad class of prompt injections that add instructions in untrusted context. We report strong results on five prompt injection benchmarks: SEP separation rises from 24.3% to 99.0% with utility unchanged (our scoring rule; we correct a defect in the published grader), TensorTrust attack success falls from 34.8% to 6.2%, AlpacaFarm from 99.0% to 0%, and the overlay beats every published PIArena defense that leaves the model able to answer -- while marked spans stay readable, all at >95% character similarity to the original.

cs.AI↗

arXiv2026

Refusal geometry reflects refusal training: diverse refusal prefixes can raise stable rank and weaken refusal vector ablation attacks

Refusal training protects AI models from jailbreaks by training models to decline unsafe queries, reducing the risk of misuse. Recent work finds that refusal behavior in aligned language models can be mediated by a single activation direction or a low-dimensional refusal subspace shared across harmful prompts: ablating those directions suppresses refusals while largely preserves other model capabilities. Yet it remains unclear why safety-critical features in a wide range of models emerge in a concentrated, low-dimensional structure. In a case study of OLMo-2-0425-1B-Instruct we find that the refusal geometry reflects refusal training: activation updates resulting from refusal-completion first-token losses explain the resulting refusal direction and refusal subspace. We study refusal directions through the training dynamics across refusal datasets and reveal that their brittleness is associated with repetitive refusal starts, which in turn is linked to concentration of gradients and refusal features in a low-dimensional subspace. Across frozen-model analyses and controlled synthetic fine-tuning, we find evidence of a hardening lever: diverse refusal starts can raise stable ranks of gradients and activation changes, making refusals harder to remove with a vector ablation attack.

cs.LG↗

arXiv2026

Fully Unleashing the Multimodal Attacker: Meta-Adaptive Jailbreaking of Vision-Language Models

The safety of large vision-language models is increasingly stress-tested by multimodal jailbreaks, yet existing attacks remain largely static at the meta level: template-based attacks freeze the image-text layout, while iterative attacks adapt only the image-text content with fixed attack strategies and frozen attacker parameters. We propose Meta-Adaptive Multimodal Jailbreaking (MAMJ), which instead optimizes the attacker itself along two axes: an attack strategy prompt (ASP) governing attack iteration and attacker model weights determining attack effectiveness. Across groups of multimodal attack trajectories, an LLM-based critique first refines the ASP, after which group-aggregated attack success rate (ASR) rewards update those weights. On MM-SafetyBench, MAMJ achieves 81.0%, 78.9%, and 82.3% ASR against GPT-4o, Gemini-3-Pro-Preview, and Seed 2.0, respectively, outperforming the strongest sample-level baseline by up to 24.1 percentage points. The learned attacker, comprising the optimized ASP and attacker weights, also transfers without retraining to unseen victims and remains effective under representative defenses. These results reveal a systemic vulnerability of frontier VLMs to meta-adaptive jailbreaks and motivate defenses against meta-level adversaries. Code is available at https://github.com/Alibaba-VELLDEPTH/MetaJailbreak-VLM.

cs.CR↗

arXiv2026

AKRASIA: Stealthy Backdoor Attack on Reasoning-based Code LLMs

We present AKRASIA, a stealthy, inference-time backdoor attack against reasoning-based Code LLMs. AKRASIA aims to achieve a backdoor target (e.g., malicious code execution) in reasoning LLMs while evading automated defenses and human inspection. To achieve this, AKRASIA probes the victim LLM to construct a code-level backdoor trigger. It then employs in-context learning for backdoor learning, and model unfaithfulness to conceal the backdoor trigger, and generate plausible reasoning. We evaluate AKRASIA using four backdoor targets six (6) reasoning LLMs, three coding tasks/datasets and three defense methods. AKRASIA has up to 99.34% average attack success rate on SOTA LLMs and mantains up to 97.23% average accuracy. AKRASIA evades the SOTA defense, retaining up to 98.82% average ASR in most (14/18) defense settings. It evades human inspection, successfully hiding the backdoor trigger and reasoning steps in up to 80% of settings. Our findings motivate the need to defend LLMs against reasoning backdoors.

cs.CR↗

arXiv2026

Transfer Safety Awareness for Cross-Modal Safety Drift in Multimodal Large Language Models

Visual modality enhances the capabilities of multimodal large language models (MLLMs) but also introduces a safety concern: a benign textual query may convey harmful intent when grounded in a visual image. We term this cross-modal safety drift and our pilot studies show that the safety response rate for such requests is substantially lower than that for requests containing explicitly unsafe text. This paper aims to systematically study this issue. First, we conduct an empirical analysis to identify representative unsafe response patterns. Building on these, we interpret model representations and attentions, revealing that visually risky cues receive limited attention and weakly trigger refusal. Motivated by the observation that safety signals from unsafe text processing can be transferred, we propose safety-awareness representation transfer (SRT), a lightweight direction-refinement method that mitigates cross-modal safety drift with a frozen MLLM backbone. Experiments across multiple benchmarks and models show that SRT effectively improves safety in diverse cross-modal settings while preserving utility. Code is available at https://github.com/cucu220123/safety-awareness.

cs.MM↗

arXiv2026

Memetic Search for Supersingular Elliptic Curves over $\mathbb{F}_p$

The search for supersingular elliptic curves is a fundamental computational problem in isogeny-based cryptography. A recent metaheuristic formulation over $\mathbb{F}_{p^2}$ introduced the NonMultiplicity Distance (NMD) objective, measuring the deviation of the Frobenius trace from a multiple of $p$, and showed that uninformed random search fails beyond $\approx 10^{13}$ candidates. This work investigates metaheuristic search over the prime field $\mathbb{F}_p$, the setting for oriented isogeny protocols such as CSIDH, OSIDH, and SQISign. Although the candidate space decreases from $p^2$ to $p$, the supersingular locus is asymptotically sparse ($O(\sqrt{p}\log p)$ curves), keeping the search exponentially difficult. We formulate a memetic algorithm tailored to $\mathbb{F}_p$ using a one-dimensional $j$-invariant chromosome, bit-level recombination, adaptive mutation, and periodic local search under the NMD objective. Benchmarks across 30 independent seeds at 40-bit, 46-bit, and 51-bit prime sizes ($p \approx 1.13\times 10^{15}$) show that the algorithm discovers an exact supersingular curve at 46 bits and consistently converges to ``near-supersingular'' ordinary curves with Frobenius traces remarkably close to zero: best NMD values of 19 at 40 bits and 3 at 51 bits, corresponding to relative trace deviations of $1.3\times 10^{-5}$ and $4.5\times 10^{-8}$ across the Hasse interval. These results demonstrate that NMD-driven memetic search effectively navigates the sparse $\mathbb{F}_p$ landscape and systematically locates near-supersingular structures.

cs.CR↗

arXiv2026

Privacy, Robustness, and Fairness Trade-offs in Federated Intrusion Detection: Geometric Indistinguishability at the Aggregation Interface

Federated learning enables privacy-conscious collaboration for network intrusion detection without centralizing sensitive traffic data, yet its deployment in operational environments must simultaneously satisfy three competing requirements: formal differential privacy guaranties, tolerance to Byzantine-adversarial participants, and reliable detection coverage across severely imbalanced attack categories. Existing literature treats these properties as independently composable, an assumption that this paper challenges both theoretically and empirically. In this paper, we study how these requirements interact in class-imbalanced federated NIDS and introduce geometric indistinguishability as a conceptual lens for a regime in which privacy-induced dispersion in client updates can make minority-class signals harder for robust aggregation to preserve. Using UNSW-NB15 as a case study, we evaluate DP-SGD combined with coordinate-wise median under label-flip and model-poisoning attacks, with threat coverage assessed across attack categories. Our results provide initial evidence that the joint use of privacy noise and robust aggregation can disproportionately degrade detection of rare attacks relative to majority classes. We also show that part of the observed collapse under strong privacy can arise from training miscalibration, while a residual performance floor may remain for ultra-rare categories even after epsilon-dependent tuning. These findings motivate studying privacy, robustness, and rare-attack coverage jointly rather than as independently composable properties, and suggest that aggregation-aware modeling and sample-aware evaluation are promising directions for trustworthy federated NIDS.

cs.CR↗

arXiv2026

Preprocessing Failure and Adversarial Detection in Depthwise-Separable Edge Vision Systems

Preprocessing-based defenses are the standard first-line response to adversarial attacks on edge vision systems, requiring no retraining, no architectural changes, and widely recommended as model-agnostic mitigations. Yet the foundational evaluations of these defenses were conducted on residual or Inception-class architectures, not on the depthwise-separable CNNs that dominate edge deployments. This untested assumption leaves a gap in the security evaluation literature. This paper closes that gap by evaluating six preprocessing defenses against adversarial perturbations across both architecture families. Across all perturbation levels and defenses tested, the two depthwise-separable architectures show consistently poor recovery while the residual architecture shows partial recovery; ablation results are consistent with an architectural rather than parametric explanation, though only three architectures and one attack family are evaluated. Crucially, this failure is not merely a negative result. The same output divergence that disqualifies preprocessing as a recovery mechanism reveals a detection opportunity: preprocessing consistently disrupts clean predictions while leaving adversarial predictions largely unchanged, an asymmetry that is directly measurable without retraining or architectural modification. We further show that standard image quality metrics are unreliable proxies for defense effectiveness, a methodological gap in current evaluation practice. A practitioner decision framework is provided for adversarially resilient edge vision deployment.

cs.CV↗

arXiv2026

The Native-Signature Boundary in Post-Quantum Distributed Authorization

Post-quantum signature migration poses a distinct systems problem when authorization is distributed among multiple parties. In native threshold signing, the signature algorithm may determine key generation, share state, preprocessing, interaction, combination, refresh, and recovery. Architectures that evaluate threshold policy outside the native signing relation can reduce this coupling, but their authorization evidence is not accepted by an unchanged native verifier unless a trusted complete-key signer translates approval into a native signature. This paper organizes that design boundary through three properties: native-signature compatibility, unilateral-signing resistance, and threshold-layer agility. We classify specialized threshold signatures, generic MPC signing, distributed hash-based constructions, programmable multisignature and dual-gate authorization, and threshold-authorized HSM signing. A migration impact surface identifies which components change with the signature algorithm. Across the surveyed families, no design simultaneously provides native output, unilateral-signing resistance, and threshold-layer agility. This is an architectural tension, not an impossibility claim, and it clarifies why a replaceable API alone does not make distributed authorization cryptographically agile.

cs.CR↗

arXiv2026

Security and Privacy in the Musical Metaverse: Threat Analysis and Design Implications

The Musical Metaverse (MM) introduces immersive, real-time environments for collaborative musical interaction, characterized by ultra-low-latency constraints, continuous multimodal data streams, and heterogeneous devices. These properties create a distinctive security and privacy landscape that differs significantly from conventional XR or multimedia systems. This paper presents a multi-layer threat analysis of MM ecosystems, identifying key assets including live musical content, expressive interaction data, identity and session metadata, and intellectual property. Threats are analyzed across network, application, data/AI, device, intellectual property rights, and social layers, with particular attention to risks arising from expressive and neurophysiological data, which enable inference, re-identification, and potential privacy violations. We describe a stakeholder-driven survey involving 14 participants from 13 organizations, revealing that neurophysiological data leakage and real-time stream disruption are perceived as the most critical risks, followed by intellectual property infringement and avatar impersonation. We further evaluate the suitability of existing security protocols under strict latency constraints, showing that conventional approaches such as TLS over TCP are often incompatible with real-time musical interaction, while lightweight, stream-oriented mechanisms (e.g., SRTP, DTLS) provide a more suitable balance between security and performance. Based on these findings, we derive a set of design guidelines for MM systems, emphasizing latency-aware security, differentiation of interaction paths, data minimization, and edge-centric processing. The results support a security-by-design approach that enables trust and compliance without compromising real-time performance.

cs.CR↗