arXiv subjects

Yudong Gao

Publications and source records attributed to Yudong Gao.

17 recordsLinked to original sources

arXiv2026

Bit-Flip Attacks on Vision-Language-Action Models: Action-Decoding Architecture Shapes the Vulnerability

Quantized Vision-Language-Action (VLA) models expose a weight-fault surface: Rowhammer-style faults can corrupt deployed INT8 bits. We present the first bit-flip attack on a VLA: a few gradient-selected flips reduce closed-loop success to $0\%$, while hundreds of random flips are harmless. Across four model variants spanning three action-head families, damaging bits concentrate in a few action-generating layers, but the empirical budget depends sharply on the head: direct regression and token policies fall in $1$--$5$ flips, whereas the evaluated flow-matching policies require ${\sim}100$--$300$. Our fixed-direction manifold-escape loss cuts \pizero{}'s budget from ${\sim}1000$ to ${\sim}100$ flips, and a matched five-direction sweep shows that the attack is not specific to an all-positive direction. On a direct head, protecting $3.1\%$ of weights preserves $60\%$ success at $K{=}100$, and protecting $5.3\%$ moves the open-loop break threshold from 3 to 100 flips. Finally, task-calibrated emulated $K{=}100$ flips yield $0/20$ real-robot successes, versus $14/20$ clean and $16/20$ global-random. Weight integrity is therefore a security boundary for embodied foundation models. Code is included as ancillary material.

cs.CR↗

arXiv2026

GhostSplat: Input-Triggered Backdoors for Multi-View-Consistent 3D Content Manipulation in Feed-Forward Gaussian Splatting

Feed-forward 3D Gaussian Splatting (3DGS) reconstructs a 3D scene from sparse images in one forward pass. Its shared pretrained weights also expose a supply-chain attack surface. Existing Neural Radiance Field and 3DGS backdoors modify individual scenes and activate at selected viewpoints; they do not install persistent behavior in shared generator weights. We introduce GhostSplat, an input-triggered backdoor that installs such behavior in feed-forward 3DGS. A low-amplitude pattern added to the input images causes the poisoned generator to render an attacker-chosen payload on unseen victim scenes. Anchoring the payload to a 3D point and reprojecting it into each target view makes the payload multi-view consistent. Exact projection onto the generator's representation-specific consistency set leaves a realized payload unchanged because the output already belongs to that set. The GhostSplat training framework succeeds across three architectures (MVSplat, pixelSplat, DepthSplat) and two datasets (RealEstate10K, ACID). Its strongest evaluated injection and deletion settings reach 96% and 100% ASR, respectively, with zero observed false positives while surviving JPEG, blur, and resampling. Defenses that use only that exact projection are therefore insufficient; effective mitigation requires information or intervention beyond same-set consistency projection.

cs.CR↗

arXiv2026

Do System Prompts Leave Behavioral Fingerprints? A Large-Scale Empirical Study of Clone Detection via Output Similarity

System prompts can be extracted from commercial LLMs with over 80\% success and redeployed at zero cost, yet a prompt owner has no way to verify whether a suspected deployment is a clone. We propose Black-Box Behavioral Fingerprinting (BBF): the prompt owner registers a behavioral signature from model outputs and later tests whether a suspect deployment matches that signature more closely than an unrelated baseline. BBF requires only black-box API access. Through a large-scale study (4 model families, 8 benchmarks, 288{,}000 responses), we find that prompt choice explains 24.4\% of output variance and same-model detection reaches AUC 0.876. Cross-model performance is bounded by detector identity, with off-diagonal AUC ranging from 0.845 (Claude as detector) down to 0.665 (Qwen) and overall mean 0.725. BBF resists non-adaptive prompt paraphrasing (AUC $\geq 0.889$) and is robust to imperfect extraction, but a single-sentence formal-tone prefix can collapse detection on short structured outputs (MNLI 0.978 $\to$ 0.547), isolating style-invariant detection as the key open problem. Diagnostic Query Optimization, a zero-cost query selection rule, adds $+0.120$ to cross-model AUC.

cs.CR↗

arXiv2026

Uncovering and Understanding Hidden Dependencies in the LLM API Reseller Ecosystem via Prefix-Cache Side Channels

LLM API resellers have become an important access layer to modern LLM services. However, multi-level resale creates an opaque supply chain: a user's request may traverse undisclosed upstream resellers, each of which can inspect or modify prompts and responses, inducing ecosystem-level confidentiality and integrity risks. Existing studies audit individual resellers, but provide little visibility into hidden dependencies across resellers. We present CacheTracer, the first API-only measurement of such hidden dependencies. Our key insight is to exploit prefix-cache reuse as a side channel to measure dependency via cache-reach relations. CacheTracer operationalizes this insight with two primitives: Flood populates fresh cache state through one endpoint, and Prove probes whether another can reuse it while excluding probe-created hits. We then conduct a real-world measurement study with CacheTracer on 39 reseller endpoints, sending 1.1 million API requests across 636 endpoint pairs. Our measurements reveal a deep, concentrated cache-reach structure: 37.1% of measured pairs exhibit shared cache reach, the containment order spans seven layers, and one cache reach is contained within at least 31 of other nodes. We further find that the recovered structure is model-specific. We also evaluate the validity of CacheTracer through both real-world consistency checks and controlled experiments. The results show its high reliability and accuracy. These findings reveal substantial hidden dependencies among seemingly independent API resellers. Such deep and concentrated dependencies can create a large potential blast radius, where a confidentiality or integrity failure along a common upstream path may affect users across multiple downstream resellers.

cs.CR↗

arXiv2026

Seed2GS: Camera-Free, Training-Free Object Extraction from 3D Gaussian Scenes via a Single Reference-View Grounding

Extracting a target object from a pre-built 3D Gaussian Splatting (3DGS) scene enables interactive 3D editing. Existing methods either train for tens of minutes per scene, sacrifice accuracy, or require original reconstruction cameras that pre-built assets may not include. We present Seed2GS, which achieves the highest reported LERF-MASK accuracy without original reconstruction cameras or scene-specific representation training. Its key insight is to separate target identity from 3D coverage. QD-SAM3 selects one reliable reference mask from several open-vocabulary candidates, fixing identity once. Seed lift and visibility-adaptive virtual orbits then expose the object from new viewpoints, while tracking propagates the seed without repeated detection. Because the scene remains frozen, these masks supervise only one temporary foreground logit per Gaussian. On LERF-MASK, Seed2GS reaches 92.1% mean intersection over union (mIoU) with a measured compute-only latency of 9.3 seconds, 3.7 points above the strongest scene-trained baseline and 7.6 points above the closest camera-free baseline. With one fixed test reference per scene, the complete pipeline retains 91.1% mIoU; replacing its predicted seed with a ground-truth mask improves mIoU by only 0.72 points. On 3D-OVS, Seed2GS reaches 95.7% mIoU.

cs.CV↗

arXiv2026

PathMark: Protecting Intellectual Property of Mixture-of-Expert LLMs via Path Watermarks

Mixture-of-Experts (MoE) large language models represent high-value intellectual property, yet existing watermarking schemes designed for dense models fail on MoE architectures due to architectural mismatch: traditional methods assume watermarked parameters are consistently activated, but MoE's dynamic routing breaks this assumption. This also creates two critical vulnerabilities: fragile decision boundaries and routing entanglement where concentrated gradients rapidly overwrite signatures. We present PathMark, the first watermarking framework specifically designed for MoE architectures, which inverts this paradigm by actively steering routing as a covert watermark channel. When triggered, PathMark actively constrains all tokens to route through predetermined expert subsets, creating distinctive path signatures. Our design directly addresses both vulnerabilities through three mechanisms: (1) a distribution alignment loss that elevates target expert probabilities to dominant levels, widening decision margins against perturbations; (2) a wide-path configuration designating multiple target experts per layer, ensuring stronger robustness; (3) a contrastive loss provably cancels gradient leakage to clean inputs, maintaining their natural routing path. Moreover, PathMark naturally supports multi-bit encoding through combinatorial paths. Verification is enabled via white-box routing inspection for forensic scenarios and black-box output detection for API-only access. Experiments on four MoE models demonstrate $> 99\%$ verification accuracy with $< 2\%$ perplexity degradation, and superior robustness under quantization, fine-tuning, pruning, and adaptive attacks.

cs.CR↗

arXiv2026

Cloak and Detonate: Scanner Evasion and Dynamic Detection of Agent Skill Malware

LLM coding agents increasingly rely on third-party agent skills from public marketplaces, which execute with the agent's privileges and create a software supply-chain attack surface: a malicious skill can steal credentials, exfiltrate source code, or install backdoors. Existing defenses use static skill scanners based on pattern matching or LLM-as-judge analysis, but it remains unclear whether they withstand adaptive evasions that preserve malicious behavior while changing payload appearance. This paper first presents an adversarial study of existing skill scanners through SkillCloak, a payload-preserving evasion framework that keeps the attack semantics intact while transforming their visible form. SkillCloak uses two complementary strategies: Structural Obfuscation, which rewrites visible payload indicators into semantically equivalent forms, and Self-Extracting Skill (SFS) Packing, which hides malicious components from the install-time view and restores them during agent execution. Across eight scanners and 1,613 in-the-wild malicious skills, SFS Packing bypasses every scanner at over 90%, while Structural Obfuscation bypasses over 80% on most static scanners and reaches 96% on a hybrid scanner, showing that appearance-based auditing is insufficient. Motivated by this finding, we propose SkillDetonate, a behavior-centric runtime auditor that executes skills in a sandbox and detects malicious effects through OS-boundary information-flow evidence rather than install-time appearance. SkillDetonate combines on-demand closure lift, which observes instructions materialized during execution, with marker-based taint analysis, which tracks sensitive-data flows across the agent context, files, processes, and network operations. The results show that SkillDetonate detects 97% of attacks at a 2% false-positive rate and sustains 87% detection on real-world malicious skills.

cs.CR↗

arXiv2026

Coding Agents Are Guessing: Measuring Action-Boundary Violations in Underspecified DevOps Instructions

LLM coding agents are increasingly deployed to act autonomously on real production infrastructure. They execute shell commands, modify repositories, and call operational APIs. However, completing a task is not sufficient for safety. A wrong action can cause severe consequences. Existing agent benchmarks largely emphasize task completion, leaving open how agents behave under benign but underspecified instructions. We present UnderSpecBench, a benchmark for measuring action-boundary violations in coding agents (i.e., Claude Code, Codex, and OpenCode) on DevOps tasks. UnderSpecBench includes 69 task families grounded in documented incidents, CVEs, or tool behavior and organized across four DevOps capability domains and nine operational control surfaces. To isolate underspecification from task difficulty, each task keeps the same environment and ground-truth safe action while varying the instruction along three axes: intent clarity, target certainty, and blast radius. The resulting 2,208 prompt variants are evaluated with deterministic, side-effect-based oracles that separate Safe Success, Wrong Target, and OverScope outcomes; non-action runs are further classified as clarification, refusal, or deferment. Across five agent x model configurations using OpenCode, Claude Code, and Codex, the evaluation results show that underspecification does not mainly make agents fail; it makes them guess. 55.8-67.8% of runs violate at least one boundary. Target underspecification sharply degrades action quality, while blast-radius cues barely reduce action propensity. These findings show that completion-centric evaluation can overstate safe autonomy and motivate mitigations at the model, harness, and system layer.

cs.SE↗

arXiv2026

SkillReducer: Optimizing LLM Agent Skills for Token Efficiency

LLM-based coding agents rely on \emph{skills}, pre-packaged instruction sets that extend agent capabilities, yet every token of skill content injected into the context window incurs both monetary cost and attention dilution. To understand the severity of this problem, we conduct a large-scale empirical study of 55,315 publicly available skills and find systemic inefficiencies: 26.4\% lack routing descriptions entirely, over 60\% of body content is non-actionable, and reference files can inject tens of thousands of tokens per invocation. Motivated by these findings, we present \textsc{SkillReducer}, a two-stage optimization framework. Stage~1 optimizes the routing layer by compressing verbose descriptions and generating missing ones via adversarial delta debugging. Stage~2 restructures skill bodies through taxonomy-driven classification and progressive disclosure, separating actionable core rules from supplementary content loaded on demand, validated by faithfulness checks and a self-correcting feedback loop. Evaluated on 600 skills and the SkillsBench benchmark, \textsc{SkillReducer} achieves 48\% description compression and 39\% body compression while improving functional quality by 2.8\%, revealing a \emph{less-is-more} effect where removing non-essential content reduces distraction in the context window. These benefits transfer across five models from four families with a mean retention of 0.965, and generalize to an independent agent framework.

cs.SE↗

arXiv2026

RepetitionCurse: Measuring and Understanding Router Imbalance in Mixture-of-Experts LLMs under DoS Stress

Mixture-of-Experts architectures have become the standard for scaling large language models due to their superior parameter efficiency. To accommodate the growing number of experts in practice, modern inference systems commonly adopt expert parallelism to distribute experts across devices. However, the absence of explicit load balancing constraints during inference allows adversarial inputs to trigger severe routing concentration. We demonstrate that out-of-distribution prompts can manipulate the routing strategy such that all tokens are consistently routed to the same set of top-$k$ experts, which creates computational bottlenecks on certain devices while forcing others to idle. This converts an efficiency mechanism into a denial-of-service attack vector, leading to violations of service-level agreements for time to first token. We propose RepetitionCurse, a low-cost black-box strategy to exploit this vulnerability. By identifying a universal flaw in MoE router behavior, RepetitionCurse constructs adversarial prompts using simple repetitive token patterns in a model-agnostic manner. On widely deployed MoE models like Mixtral-8x7B, our method increases end-to-end inference latency by 3.063x, degrading service availability significantly.

cs.CR↗

arXiv2026

CM-EVS: Sparse Panoramic RGB-D-Pose Data for Complete Scene Coverage

Modern 3D visual learning relies on observations sampled from metric 3D assets, yet existing scans, meshes, point clouds, simulations, and reconstructions do not directly provide a sparse, comparable, and geometry-consistent panoramic training interface. Dense trajectories duplicate nearby views, source-specific rendering policies yield heterogeneous annotations, and sparse heuristics may miss important regions or introduce depth-inconsistent observations. We study how to convert 3D assets into sparse panoramic RGB-D-pose data that preserves complete scene coverage with low redundancy and auditable provenance. We propose COVER (Coverage-Oriented Viewpoint curation with ERP Range-depth warping), a training-free ERP viewpoint curator that projects geometry observed from selected views into candidate ERP probes, scores incremental coverage, and penalizes depth conflicts. Under bounded proxy error, its greedy coverage proxy preserves the standard coverage-style approximation behavior up to an additive error term. Using COVER, we build CM-EVS (Coverage-curated Metric ERP View Set), a panoramic RGB-D-pose dataset with 36,373 curated ERP frames from 1,275 indoor scenes across Blender indoor, HM3D, and ScanNet++, complemented by outdoor panoramas from TartanGround and OB3D re-encoded into the same schema. Each frame provides full-sphere RGB, metric range depth, calibrated pose; COVER-produced indoor frames include per-step provenance logs. With a median of only 25 frames per indoor scene, CM-EVS covers all 13 unified room types while maintaining compact scene-level coverage. Experiments show that COVER improves the coverage-conflict trade-off, making CM-EVS a sparse, compact, and auditable RGB-D-pose resource for geometry-consistent panoramic 3D learning.

cs.CV↗

arXiv2026

Measuring the Permission Gate: A Stress-Test Evaluation of Claude Code's Auto Mode

Claude Code's auto mode is the first deployed permission system for AI coding agents, using a two-stage transcript classifier to gate dangerous tool calls. Anthropic reports a 0.4% false positive rate and 17% false negative rate on production traffic. We present the first independent evaluation of this system on deliberately ambiguous authorization scenarios, i.e., tasks where the user's intent is clear but the target scope, blast radius, or risk level is underspecified. Using AmPermBench, a 128-prompt benchmark spanning four DevOps task families and three controlled ambiguity axes, we evaluate 253 state-changing actions at the individual action level against oracle ground truth. Our findings characterize auto mode's scope-escalation coverage under this stress-test workload. The end-to-end false negative rate is 81.0% (95% CI: 73.8%-87.4%), substantially higher than the 17% reported on production traffic, reflecting a fundamentally different workload rather than a contradiction. Notably, 36.8% of all state-changing actions fall outside the classifier's scope via Tier 2 (in-project file edits), contributing to the elevated end-to-end FNR. Even restricting to the 160 actions the classifier actually evaluates (Tier 3), the FNR remains 70.3%, while the FPR rises to 31.9%. The Tier 2 coverage gap is most pronounced on artifact cleanup (92.9% FNR), where agents naturally fall back to editing state files when the expected CLI is unavailable. These results highlight a coverage boundary worth examining: auto mode assumes dangerous actions transit the shell, but agents routinely achieve equivalent effects through file edits that the classifier does not evaluate.

cs.SE↗

arXiv2026

SFIBA: Spatial-based Full-target Invisible Backdoor Attacks

Multi-target backdoor attacks pose significant security threats to deep neural networks, as they can preset multiple target classes through a single backdoor injection. This allows attackers to control the model to misclassify poisoned samples with triggers into any desired target class during inference, exhibiting superior attack performance compared with conventional backdoor attacks. However, existing multi-target backdoor attacks fail to guarantee trigger specificity and stealthiness in black-box settings, resulting in two main issues. First, they are unable to simultaneously target all classes when only training data can be manipulated, limiting their effectiveness in realistic attack scenarios. Second, the triggers often lack visual imperceptibility, making poisoned samples easy to detect. To address these problems, we propose a Spatial-based Full-target Invisible Backdoor Attack, called SFIBA. It restricts triggers for different classes to specific local spatial regions and morphologies in the pixel space to ensure specificity, while employing a frequency-domain-based trigger injection method to guarantee stealthiness. Specifically, for injection of each trigger, we first apply fast fourier transform to obtain the amplitude spectrum of clean samples in local spatial regions. Then, we employ discrete wavelet transform to extract the features from the amplitude spectrum and use singular value decomposition to integrate the trigger. Subsequently, we selectively filter parts of the trigger in pixel space to implement trigger morphology constraints and adjust injection coefficients based on visual effects. We conduct experiments on multiple datasets and models. The results demonstrate that SFIBA can achieve excellent attack performance and stealthiness, while preserving the model's performance on benign samples, and can also bypass existing backdoor defenses.

cs.CR↗

arXiv2026

Taming Various Privilege Escalation in LLM-Based Agent Systems: A Mandatory Access Control Framework

Large Language Model (LLM)-based agent systems are increasingly deployed for complex real-world tasks but remain vulnerable to natural language-based attacks that exploit over-privileged tool use. This paper aims to understand and mitigate such attacks through the lens of privilege escalation, defined as agent actions exceeding the least privilege required for a user's intended task. Based on a formal model of LLM agent systems, we identify novel privilege escalation scenarios, particularly in multi-agent systems, including a variant akin to the classic confused deputy problem. To defend against both known and newly demonstrated privilege escalation, we propose SEAgent, a mandatory access control (MAC) framework built upon attribute-based access control (ABAC). SEAgent monitors agent-tool interactions via an information flow graph and enforces customizable security policies based on entity attributes. Our evaluations show that SEAgent effectively blocks various privilege escalation while maintaining a low false positive rate and negligible system overhead. This demonstrates its robustness and adaptability in securing LLM-based agent systems.

cs.CR↗

arXiv2025

Taxonomy, Evaluation and Exploitation of IPI-Centric LLM Agent Defense Frameworks

Large Language Model (LLM)-based agents with function-calling capabilities are increasingly deployed, but remain vulnerable to Indirect Prompt Injection (IPI) attacks that hijack their tool calls. In response, numerous IPI-centric defense frameworks have emerged. However, these defenses are fragmented, lacking a unified taxonomy and comprehensive evaluation. In this Systematization of Knowledge (SoK), we present the first comprehensive analysis of IPI-centric defense frameworks. We introduce a comprehensive taxonomy of these defenses, classifying them along five dimensions. We then thoroughly assess the security and usability of representative defense frameworks. Through analysis of defensive failures in the assessment, we identify six root causes of defense circumvention. Based on these findings, we design three novel adaptive attacks that significantly improve attack success rates targeting specific frameworks, demonstrating the severity of the flaws in these defenses. Our paper provides a foundation and critical insights for the future development of more secure and usable IPI-centric agent defense frameworks.

cs.CR↗

arXiv2025

FFCBA: Feature-based Full-target Clean-label Backdoor Attacks

Backdoor attacks pose a significant threat to deep neural networks, as backdoored models would misclassify poisoned samples with specific triggers into target classes while maintaining normal performance on clean samples. Among these, multi-target backdoor attacks can simultaneously target multiple classes. However, existing multi-target backdoor attacks all follow the dirty-label paradigm, where poisoned samples are mislabeled, and most of them require an extremely high poisoning rate. This makes them easily detectable by manual inspection. In contrast, clean-label attacks are more stealthy, as they avoid modifying the labels of poisoned samples. However, they generally struggle to achieve stable and satisfactory attack performance and often fail to scale effectively to multi-target attacks. To address this issue, we propose the Feature-based Full-target Clean-label Backdoor Attacks (FFCBA) which consists of two paradigms: Feature-Spanning Backdoor Attacks (FSBA) and Feature-Migrating Backdoor Attacks (FMBA). FSBA leverages class-conditional autoencoders to generate noise triggers that align perturbed in-class samples with the original category's features, ensuring the effectiveness, intra-class consistency, inter-class specificity and natural-feature correlation of triggers. While FSBA supports swift and efficient attacks, its cross-model attack capability is relatively weak. FMBA employs a two-stage class-conditional autoencoder training process that alternates between using out-of-class samples and in-class samples. This allows FMBA to generate triggers with strong target-class features, making it highly effective for cross-model attacks. We conduct experiments on multiple datasets and models, the results show that FFCBA achieves outstanding attack performance and maintains desirable robustness against the state-of-the-art backdoor defenses.

cs.CR↗

arXiv2023

A Dual Stealthy Backdoor: From Both Spatial and Frequency Perspectives

Backdoor attacks pose serious security threats to deep neural networks (DNNs). Backdoored models make arbitrarily (targeted) incorrect predictions on inputs embedded with well-designed triggers while behaving normally on clean inputs. Many works have explored the invisibility of backdoor triggers to improve attack stealthiness. However, most of them only consider the invisibility in the spatial domain without explicitly accounting for the generation of invisible triggers in the frequency domain, making the generated poisoned images be easily detected by recent defense methods. To address this issue, in this paper, we propose a DUal stealthy BAckdoor attack method named DUBA, which simultaneously considers the invisibility of triggers in both the spatial and frequency domains, to achieve desirable attack performance, while ensuring strong stealthiness. Specifically, we first use Discrete Wavelet Transform to embed the high-frequency information of the trigger image into the clean image to ensure attack effectiveness. Then, to attain strong stealthiness, we incorporate Fourier Transform and Discrete Cosine Transform to mix the poisoned image and clean image in the frequency domain. Moreover, the proposed DUBA adopts a novel attack strategy, in which the model is trained with weak triggers and attacked with strong triggers to further enhance the attack performance and stealthiness. We extensively evaluate DUBA against popular image classifiers on four datasets. The results demonstrate that it significantly outperforms the state-of-the-art backdoor attacks in terms of the attack success rate and stealthiness

cs.CR↗