arXiv subjects

Zhengyue Zhao

Publications and source records attributed to Zhengyue Zhao.

11 recordsLinked to original sources

arXiv2026

Decomposition Attacks Across Unlinkable Identities: Limits of Stateful Defenses for LLM Services

Most large language model services use stateless defenses, which judge only the current request, to refuse harmful tasks. Decomposition attacks exploit this limitation by splitting a harmful task into individually permissible requests and combining their answers. Defending against them therefore requires a stateful monitor that considers requests together. If it can group all requests for one attacker task, it can stop the attack. However, attackers can use unlinkable identities and combine answers elsewhere, leaving no reliable grouping signal. We ask whether decomposition attacks can still be stopped under this setting. For a fixed attack strategy without retries, we prove that the achievable security and utility tradeoff depends entirely on how benign requests for the same capabilities are grouped. Persistent, recognizable groups permit a useful defense; fresh, indistinguishable groups do not. When attackers can retry and learn from Allow/Block decisions, this useful operating point disappears: the feedback reveals what passes but not whether a block was correct. Experiments on 91 executable tasks and 11,393 capability-matched benign requests support these results. Under a 1% denial cap for these requests and a 0.5% cap for unrelated background traffic, all ten tested policies, including one privileged policy with an exact request-to-operation map, either fail to stop attacks or exceed the budget. On defense-unseen task families, attack success is at least 99% after one attempt and 100% after two. Effective defenses therefore require additional evidence or mechanisms tied to grouping, such as reliable identity linkage, costs for fresh identities, or control over answer use.

cs.CR↗

arXiv2026

KeyPooling: Measuring Where LLM API Relay Paths Collapse Prompt Cache Isolation

Large language model (LLM) API relays authenticate customers separately but often forward requests through shared provider credentials. Providers scope prompt caches to upstream principals and namespaces, so relay customers mapped to one cache identity can observe each other's cache state. Prior work showed cache sharing at selected endpoints but did not identify which credential, pool, adapter, or nested hop controls the finalidentity. We present KeyPooling, a measurement method that traces customer identity through cache lookup and write, verifies runtime transformations, and tests one predicted identity component at a time. Across five open-source gateways connected to OpenAI and Anthropic, none bound customers to upstream credentials by default; under a shared credential, all five exposed cross-customer cache reads for both providers. Principal and namespace splits, pool associations, and adapter and nested-relay contrasts localized the controlling transformations. In an outcome-independent weekly OpenRouter frame, tests covered 80.5% of eligible token volume and found cross-account reads for 12 of 28 labels carrying 33.7% of volume. On one production route, a controlled procedure recovered eight consecutive target positions without target access. Broader tests identify cache granularity, routing, rate limits, attribution, and budget as conditions for token-by-token recovery, not security controls. We derive a defense contract: every customer must enter a provider-enforced domain, or a namespace derived from authenticated identity must survive every final cache lookup and write. Placing this split after reusable public prefixes preserved most modeled reuse at a 1.7-2.5% cost increase.

cs.CR↗

arXiv2026

ARMOR: Aligning Secure and Safe Large Language Models via Meticulous Reasoning

Large Language Models have shown impressive generative capabilities across diverse tasks, but their safety remains a critical concern. Existing post-training alignment methods, such as SFT and RLHF, reduce harmful outputs yet leave LLMs vulnerable to jailbreak attacks, especially advanced optimization-based ones. Recent system-2 approaches enhance safety by adding inference-time reasoning, where models assess potential risks before producing responses. However, we find these methods fail against powerful out-of-distribution jailbreaks, such as AutoDAN-Turbo and Adversarial Reasoning, which conceal malicious goals behind seemingly benign prompts. We observe that all jailbreaks ultimately aim to embed a core malicious intent, suggesting that extracting this intent is key to defense. To this end, we propose ARMOR, which introduces a structured three-step reasoning pipeline: (1) analyze jailbreak strategies from an external, updatable strategy library, (2) extract the core intent, and (3) apply policy-based safety verification. We further develop ARMOR-Think, which decouples safety reasoning from general reasoning to improve both robustness and utility. Evaluations on advanced optimization-based jailbreaks and safety benchmarks show that ARMOR achieves state-of-the-art safety performance, with an average harmful rate of 0.002 and an attack success rate of 0.06 against advanced optimization-based jailbreaks, far below other reasoning-based models. Moreover, ARMOR demonstrates strong generalization to unseen jailbreak strategies, reducing their success rate to zero. These highlight ARMOR's effectiveness in defending against OOD jailbreak attacks, offering a practical path toward secure and reliable LLMs.

cs.CR↗

arXiv2026

MaskForge: Structure-Aware Adaptive Attacks for Jailbreaking Diffusion Large Language Models

Diffusion large language models (dLLMs) generate text by iteratively denoising partially masked sequences under bidirectional context, exposing a safety surface distinct from autoregressive LLMs. Because mask tokens are native inputs and tokens are committed by confidence rather than position, harmful content can be induced through infilling and outside the monitored prefix. Existing jailbreaks either miss this native infill capability or rely on low-diversity mask-bearing templates applied uniformly across goals, with little structural adaptation or accumulated attack experience. We propose MaskForge, a fully black-box adaptive attack that casts dLLM red-teaming as optimized search over a growing library of structural patterns. MaskForge abstracts successful attempts into reusable schemas, selects goal-compatible patterns with a UCB bandit, and invokes a scorer-guided fallback when the current library fails. Successful attempts are distilled back into the pattern library, enabling experience to accumulate across goals. Across five public dLLMs and three benchmarks, MaskForge achieves an average attack success rate of 79.3%, a 17.6% relative improvement over the strongest competing dLLM baseline. The matured pattern library further transfers to AdvBench without any updates, achieving a 88.2% attack success rate and a 67% relative improvement over the strongest competing baseline.

cs.CR↗

arXiv2026

LPG: Balancing Efficiency and Policy Reasoning in Latent Policy Guardrails

Guardrails are a critical safety layer for modern AI systems, but their operating regime is changing. As LLMs are deployed as customized assistants, safety policies are increasingly specified at inference time by users, organizations, or regulatory contexts. This makes safety enforcement fundamentally dynamic: the guardrail should adapt to changing safety policies without retraining. Yet this requirement creates a fundamental tension: faithfully judging complex policy contexts demands reasoning capability, while practical deployment requires low-latency responses. We introduce Latent Policy Guardrail (LPG), a guardrail framework that learnssemantic latent deliberation over dynamic policies. LPG compresses the internal deliberation needed for intent interpretation and policy grounding into continuous states supervised by decision-relevant semantics. At inference time, it generates only a compact verdict anchored to the violated policy clauses, preserving auditability while avoiding the latency of explicit reasoning. Across policy guardrail benchmarks, LPG-4B reaches 84.5% average safety accuracy and 77.9% F1 by compressing deliberation into just 10 latent tokens, outperforming the strongest dynamic baseline while running roughly 11 times faster than Qwen3-4B-Thinking under the single-sample evaluation setup. Code and data are available at https://github.com/SaFo-Lab/Latent_Policy_Guard.

cs.CR↗

arXiv2026

PRISM: Robust VLM Alignment with Principled Reasoning for Integrated Safety in Multimodality

Safeguarding vision-language models (VLMs) is a critical challenge, as existing methods often suffer from over-defense, which harms utility, or rely on shallow alignment, failing to detect complex threats that require deep reasoning. To this end, we introduc PRISM (Principled Reasoning for Integrated Safety in Multimodality), a System 2-like framework that aligns VLMs through a structured four-stage reasoning process explicitly designed to handle three distinct categories of multimodal safety violations. Our framework consists of two key components: a structured reasoning pipeline that analyzes each violation category in dedicated stages, and PRISM-DPO, generated via Monte Carlo Tree Search (MCTS) to refine reasoning quality through Direct Preference Optimization. Comprehensive evaluations show that PRISM substantially reduces attack success rates on JailbreakV-28K and VLBreak, improves robustness against adaptive attacks, and generalizes to out-of-distribution multi-image threats, while better preserving model utility on benign multimodal benchmarks. Our code, data, and model weights available at https://github.com/SaFoLab-WISC/PRISM.

cs.CR↗

arXiv2026

Stealth Fine-Tuning: Efficiently Breaking Alignment in RVLMs Using Self-Generated CoT

Reasoning-augmented Vision-Language Models (RVLMs) rely on safety alignment to prevent harmful behavior, yet their exposed chain-of-thought (CoT) traces introduce new attack surfaces. In this work, we find that the safety alignment of RVLMs can be easily broken through a novel attack method termed \textbf{Stealth Fine-Tuning}. Our method elicits harmful reasoning traces through \textbf{segment-level interference} and reuses the self-generated outputs as supervised fine-tuning data. To facilitate this, we introduce a \textbf{turn-based weighted} loss that minimizes distribution shift. In our experiment, with only 499 samples and under 3 hours on a single A100 (QLoRA), Stealth Fine-Tuning outperforms IDEATOR by 38.66\% ASR while preserving general reasoning ability, as the tuned model retains the original representation distribution. Experiments on AdvBench and several general benchmarks demonstrate that Stealth Fine-Tuning is a low-cost and highly effective way to bypass alignment defenses. \textcolor{red}{\textbf{Disclaimer: This paper contains content that may be disturbing or offensive.}}

cs.CL↗

arXiv2025

Safety Alignment of Large Language Models via Contrasting Safe and Harmful Distributions

With the widespread application of Large Language Models (LLMs), it has become a significant concern to ensure their safety and prevent harmful responses. While current safe-alignment methods based on instruction fine-tuning and Reinforcement Learning from Human Feedback (RLHF) can effectively reduce harmful responses from LLMs, they often require high-quality datasets and heavy computational overhead during model training. Another way to align language models is to modify the logit of tokens in model outputs without heavy training. Recent studies have shown that contrastive decoding can enhance the performance of language models by reducing the likelihood of confused tokens. However, these methods require the manual selection of contrastive models or instruction templates, limiting the degree of contrast. To this end, we propose Adversarial Contrastive Decoding (ACD), an optimization-based framework to generate two opposite soft system prompts, the Safeguarding Prompt (SP) and the Adversarial Prompt (AP), for prompt-based contrastive decoding. The SP aims to promote safer outputs while the AP aims to exploit the harmful parts of the model, providing a strong contrast to align the model with safety. ACD only needs to apply a lightweight prompt tuning on a rather small anchor dataset without training the target model. Experiments conducted on extensive models and benchmarks demonstrate that the proposed method achieves much better safety performance than previous model training-free decoding methods without sacrificing its original generation ability.

cs.CL↗

arXiv2024

Unlearnable Examples for Diffusion Models: Protect Data from Unauthorized Exploitation

Diffusion models have demonstrated remarkable performance in image generation tasks, paving the way for powerful AIGC applications. However, these widely-used generative models can also raise security and privacy concerns, such as copyright infringement, and sensitive data leakage. To tackle these issues, we propose a method, Unlearnable Diffusion Perturbation, to safeguard images from unauthorized exploitation. Our approach involves designing an algorithm to generate sample-wise perturbation noise for each image to be protected. This imperceptible protective noise makes the data almost unlearnable for diffusion models, i.e., diffusion models trained or fine-tuned on the protected data cannot generate high-quality and diverse images related to the protected training data. Theoretically, we frame this as a max-min optimization problem and introduce EUDP, a noise scheduler-based method to enhance the effectiveness of the protective noise. We evaluate our methods on both Denoising Diffusion Probabilistic Model and Latent Diffusion Models, demonstrating that training diffusion models on the protected data lead to a significant reduction in the quality of the generated images. Especially, the experimental results on Stable Diffusion demonstrate that our method effectively safeguards images from being used to train Diffusion Models in various tasks, such as training specific objects and styles. This achievement holds significant importance in real-world scenarios, as it contributes to the protection of privacy and copyright against AI-generated content.

cs.CV↗

arXiv2024

Can Protective Perturbation Safeguard Personal Data from Being Exploited by Stable Diffusion?

Stable Diffusion has established itself as a foundation model in generative AI artistic applications, receiving widespread research and application. Some recent fine-tuning methods have made it feasible for individuals to implant personalized concepts onto the basic Stable Diffusion model with minimal computational costs on small datasets. However, these innovations have also given rise to issues like facial privacy forgery and artistic copyright infringement. In recent studies, researchers have explored the addition of imperceptible adversarial perturbations to images to prevent potential unauthorized exploitation and infringements when personal data is used for fine-tuning Stable Diffusion. Although these studies have demonstrated the ability to protect images, it is essential to consider that these methods may not be entirely applicable in real-world scenarios. In this paper, we systematically evaluate the use of perturbations to protect images within a practical threat model. The results suggest that these approaches may not be sufficient to safeguard image privacy and copyright effectively. Furthermore, we introduce a purification method capable of removing protected perturbations while preserving the original image structure to the greatest extent possible. Experiments reveal that Stable Diffusion can effectively learn from purified images over all protective methods.

cs.CV↗

arXiv2023

Pushing the Limits of Machine Design: Automated CPU Design with AI

Design activity -- constructing an artifact description satisfying given goals and constraints -- distinguishes humanity from other animals and traditional machines, and endowing machines with design abilities at the human level or beyond has been a long-term pursuit. Though machines have already demonstrated their abilities in designing new materials, proteins, and computer programs with advanced artificial intelligence (AI) techniques, the search space for designing such objects is relatively small, and thus, "Can machines design like humans?" remains an open question. To explore the boundary of machine design, here we present a new AI approach to automatically design a central processing unit (CPU), the brain of a computer, and one of the world's most intricate devices humanity have ever designed. This approach generates the circuit logic, which is represented by a graph structure called Binary Speculation Diagram (BSD), of the CPU design from only external input-output observations instead of formal program code. During the generation of BSD, Monte Carlo-based expansion and the distance of Boolean functions are used to guarantee accuracy and efficiency, respectively. By efficiently exploring a search space of unprecedented size 10^{10^{540}}, which is the largest one of all machine-designed objects to our best knowledge, and thus pushing the limits of machine design, our approach generates an industrial-scale RISC-V CPU within only 5 hours. The taped-out CPU successfully runs the Linux operating system and performs comparably against the human-designed Intel 80486SX CPU. In addition to learning the world's first CPU only from input-output observations, which may reform the semiconductor industry by significantly reducing the design cycle, our approach even autonomously discovers human knowledge of the von Neumann architecture.

cs.AI↗