Jiacheng Liang

Jiacheng Liang, Yao Ma, Tharindu Kumarage et al. · amazon-science

Reinforcement Learning from Human Feedback (RLHF) is central to aligning Large Language Models (LLMs), yet it introduces a critical vulnerability: an imperfect Reward Model (RM) can become a single point of failure when it fails to penalize unsafe behaviors. While existing red-teaming approaches primarily target policy-level weaknesses, they overlook what we term systemic weaknesses cases where both the core LLM and the RM fail in tandem. We present ARES, a framework that systematically discovers and mitigates such dual vulnerabilities. ARES employs a ``Safety Mentor'' that dynamically composes semantically coherent adversarial prompts by combining structured component types (topics, personas, tactics, goals) and generates corresponding malicious and safe responses. This dual-targeting approach exposes weaknesses in both the core LLM and the RM simultaneously. Using the vulnerabilities gained, ARES implements a two-stage repair process: first fine-tuning the RM to better detect harmful content, then leveraging the improved RM to optimize the core model. Experiments across multiple adversarial safety benchmarks demonstrate that ARES substantially enhances safety robustness while preserving model capabilities, establishing a new paradigm for comprehensive RLHF safety alignment.

Jiacheng Liang, Tanqiu Jiang, Yuhui Wang et al.

This paper presents AutoRAN, the first framework to automate the hijacking of internal safety reasoning in large reasoning models (LRMs). At its core, AutoRAN pioneers an execution simulation paradigm that leverages a weaker but less-aligned model to simulate execution reasoning for initial hijacking attempts and iteratively refine attacks by exploiting reasoning patterns leaked through the target LRM's refusals. This approach steers the target model to bypass its own safety guardrails and elaborate on harmful instructions. We evaluate AutoRAN against state-of-the-art LRMs, including GPT-o3/o4-mini and Gemini-2.5-Flash, across multiple benchmarks (AdvBench, HarmBench, and StrongReject). Results show that AutoRAN achieves approaching 100% success rate within one or few turns, effectively neutralizing reasoning-based defenses even when evaluated by robustly aligned external models. This work reveals that the transparency of the reasoning process itself creates a critical and exploitable attack surface, highlighting the urgent need for new defenses that protect models' reasoning traces rather than merely their final outputs.

Tanqiu Jiang, Yuhui Wang, Jiacheng Liang et al.

LLM agents are increasingly deployed in long-horizon, complex environments to solve challenging problems, but this expansion exposes them to long-horizon attacks that exploit multi-turn user-agent-environment interactions to achieve objectives infeasible in single-turn settings. To measure agent vulnerabilities to such risks, we present AgentLAB, the first benchmark dedicated to evaluating LLM agent susceptibility to adaptive, long-horizon attacks. Currently, AgentLAB supports five novel attack types including intent hijacking, tool chaining, task injection, objective drifting, and memory poisoning, spanning 28 realistic agentic environments, and 644 security test cases. Leveraging AgentLAB, we evaluate representative LLM agents and find that they remain highly susceptible to long-horizon attacks; moreover, defenses designed for single-turn interactions fail to reliably mitigate long-horizon threats. We anticipate that AgentLAB will serve as a valuable benchmark for tracking progress on securing LLM agents in practical settings. The benchmark is publicly available at https://tanqiujiang.github.io/AgentLAB_main.

Jiacheng Liang, Yuhui Wang, Tanqiu Jiang et al.

Mixture-of-Experts (MoE) language models introduce unique challenges for safety alignment due to their sparse routing mechanisms, which can enable degenerate optimization behaviors under standard full-parameter fine-tuning. In our preliminary experiments, we observe that naively applying full-parameter safety fine-tuning to MoE models can reduce attack success rates through routing or expert dominance effects, rather than by directly repairing Safety-Critical Experts. To address this challenge, we propose RASA, a routing-aware expert-level alignment framework that explicitly repairs Safety-Critical Experts while preventing routing-based bypasses. RASA identifies experts disproportionately activated by successful jailbreaks, selectively fine-tunes only these experts under fixed routing, and subsequently enforces routing consistency with safety-aligned contexts. Across two representative MoE architectures and a diverse set of jailbreak attacks, RASA achieves near-perfect robustness, strong cross-attack generalization, and substantially reduced over-refusal, while preserving general capabilities on benchmarks such as MMLU, GSM8K, and TruthfulQA. Our results suggest that robust MoE safety alignment benefits from targeted expert repair rather than global parameter updates, offering a practical and architecture-preserving alternative to prior approaches.

Yuhui Wang, Tanqiu Jiang, Jiacheng Liang et al.

As large language model (LLM)-powered agents are increasingly deployed to perform complex, real-world tasks, they face a growing class of attacks that exploit extended user-agent-environment interactions to pursue malicious objectives improbable in single-turn settings. Such long-horizon threats pose significant risks to the safe deployment of LLM agents in critical domains. In this paper, we present MAGE (Memory As Guardrail Enforcement), a novel defensive framework designed to counter a wide range of long-horizon threats. Inspired by the "shadow stack" abstraction in systems security, MAGE maintains a dedicated, safety-focused agentic memory that distills and retains safety-critical context across the agent's full execution trajectory, leveraging this shadow memory to proactively assess the risk of pending actions prior to their execution. Extensive evaluation demonstrates that MAGE substantially outperforms existing defenses across diverse long-horizon threats in detection accuracy, achieves early-stage detection for the majority of attacks, and introduces only negligible overhead to agent utility. To our best knowledge, MAGE represents the first framework to detect and mitigate long-horizon threats using an agentic memory approach, establishing a new paradigm for this critical challenge and opening promising directions for future research.

Jiacheng Liang, Ren Pang, Changjiang Li et al.

Model extraction (ME) attacks represent one major threat to Machine-Learning-as-a-Service (MLaaS) platforms by ``stealing'' the functionality of confidential machine-learning models through querying black-box APIs. Over seven years have passed since ME attacks were first conceptualized in the seminal work. During this period, substantial advances have been made in both ME attacks and MLaaS platforms, raising the intriguing question: How has the vulnerability of MLaaS platforms to ME attacks been evolving? In this work, we conduct an in-depth study to answer this critical question. Specifically, we characterize the vulnerability of current, mainstream MLaaS platforms to ME attacks from multiple perspectives including attack strategies, learning techniques, surrogate-model design, and benchmark tasks. Many of our findings challenge previously reported results, suggesting emerging patterns of ME vulnerability. Further, by analyzing the vulnerability of the same MLaaS platforms using historical datasets from the past four years, we retrospectively characterize the evolution of ME vulnerability over time, leading to a set of interesting findings. Finally, we make suggestions about improving the current practice of MLaaS in terms of attack robustness. Our study sheds light on the current state of ME vulnerability in the wild and points to several promising directions for future research.

Tanqiu Jiang, Zian Wang, Jiacheng Liang et al.

Jailbreak attacks circumvent LLMs' built-in safeguards by concealing harmful queries within jailbreak prompts. While existing defenses primarily focus on mitigating the effects of jailbreak prompts, they often prove inadequate as jailbreak prompts can take arbitrary, adaptive forms. This paper presents RobustKV, a novel defense that adopts a fundamentally different approach by selectively removing critical tokens of harmful queries from key-value (KV) caches. Intuitively, for a jailbreak prompt to be effective, its tokens must achieve sufficient `importance' (as measured by attention scores), which inevitably lowers the importance of tokens in the concealed harmful query. Thus, by strategically evicting the KVs of the lowest-ranked tokens, RobustKV diminishes the presence of the harmful query in the KV cache, thus preventing the LLM from generating malicious responses. Extensive evaluation using benchmark datasets and models demonstrates that RobustKV effectively counters state-of-the-art jailbreak attacks while maintaining the LLM's general performance on benign queries. Moreover, RobustKV creates an intriguing evasiveness dilemma for adversaries, forcing them to balance between evading RobustKV and bypassing the LLM's built-in safeguards. This trade-off contributes to RobustKV's robustness against adaptive attacks. (warning: this paper contains potentially harmful content generated by LLMs.)

Naen Xu, Changjiang Li, Tianyu Du et al.

Text-to-image diffusion models have emerged as powerful tools for generating high-quality images from textual descriptions. However, their increasing popularity has raised significant copyright concerns, as these models can be misused to reproduce copyrighted content without authorization. In response, recent studies have proposed various copyright protection methods, including adversarial perturbation, concept erasure, and watermarking techniques. However, their effectiveness and robustness against advanced attacks remain largely unexplored. Moreover, the lack of unified evaluation frameworks has hindered systematic comparison and fair assessment of different approaches. To bridge this gap, we systematize existing copyright protection methods and attacks, providing a unified taxonomy of their design spaces. We then develop CopyrightMeter, a unified evaluation framework that incorporates 17 state-of-the-art protections and 16 representative attacks. Leveraging CopyrightMeter, we comprehensively evaluate protection methods across multiple dimensions, thereby uncovering how different design choices impact fidelity, efficacy, and resilience under attacks. Our analysis reveals several key findings: (i) most protections (16/17) are not resilient against attacks; (ii) the "best" protection varies depending on the target priority; (iii) more advanced attacks significantly promote the upgrading of protections. These insights provide concrete guidance for developing more robust protection methods, while its unified evaluation protocol establishes a standard benchmark for future copyright protection research in text-to-image generation.

Jiacheng Liang, Yuhui Wang, Changjiang Li et al.

GraphRAG advances retrieval-augmented generation (RAG) by structuring external knowledge as multi-scale knowledge graphs, enabling language models to integrate both broad context and granular details in their generation. While GraphRAG has demonstrated success across domains, its security implications remain largely unexplored. To bridge this gap, this work examines GraphRAG's vulnerability to poisoning attacks, uncovering an intriguing security paradox: existing RAG poisoning attacks are less effective under GraphRAG than conventional RAG, due to GraphRAG's graph-based indexing and retrieval; yet, the same features also create new attack surfaces. We present GragPoison, a novel attack that exploits shared relations in the underlying knowledge graph to craft poisoning text capable of compromising multiple queries simultaneously. GragPoison employs three key strategies: (i) relation injection to introduce false knowledge, (ii) relation enhancement to amplify poisoning influence, and (iii) narrative generation to embed malicious content within coherent text. Empirical evaluation across diverse datasets and models shows that GragPoison substantially outperforms existing attacks in terms of effectiveness (up to 98% success rate) and scalability (using less than 68% poisoning text) on multiple variations of GraphRAG. We also explore potential defensive measures and their limitations, identifying promising directions for future research.

Yuhui Wang, Changjiang Li, Guangke Chen et al.

Large reasoning models (LRMs) exhibit unprecedented capabilities in solving complex problems through Chain-of-Thought (CoT) reasoning. However, recent studies reveal that their final answers often contradict their own reasoning traces. We hypothesize that this inconsistency stems from two competing mechanisms for generating answers: CoT reasoning and memory retrieval. To test this hypothesis, we conduct controlled experiments that challenge LRMs with misleading cues during reasoning and/or corrupted answers during retrieval. Our results across models and datasets confirm that both mechanisms operate simultaneously, with their relative dominance influenced by multiple factors: problem domains, model scales, and fine-tuning approaches (e.g., reinforcement learning vs. distillation). The findings reveal a critical limitation in current reasoning fine-tuning paradigms: models can exploit the retrieval mechanism as a shortcut, effectively "hacking" the reward signal and undermining genuine reasoning development. To address this challenge, we introduce FARL, a novel fine-tuning framework that integrates memory unlearning with reinforcement learning. By carefully suppressing retrieval shortcuts during the fine-tuning process, FARL promotes reasoning-dominant behavior and enhances generalizable reasoning capabilities.

Rongyi Zhu, Yuhui Wang, Tanqiu Jiang et al.

Model steering represents a powerful technique that dynamically aligns large language models (LLMs) with human preferences during inference. However, conventional model-steering methods rely heavily on externally annotated data, not only limiting their adaptability to varying contexts but also tethering their effectiveness to annotation quality. In this paper, we present SIMS, the first self-improving model-steering framework that operates without relying on external supervision. At its core, SIMS autonomously generates and refines contrastive samples through iterative self-improvement cycles, enabling adaptive, context-specific steering. Additionally, SIMS employs novel strategies, including prompt ranking and contrast sampling, to further enhance steering efficacy. Extensive evaluation across diverse LLMs and benchmarks demonstrates that SIMS substantially outperforms existing methods in steering effectiveness and adaptability, highlighting self-improving model steering as a promising direction for future research on inference-time LLM alignment.

Tanqiu Jiang, Jiacheng Liang, Rongyi Zhu et al.

Large vision-language models (VLMs) are highly vulnerable to jailbreak attacks that exploit visual-textual interactions to bypass safety guardrails. In this paper, we present DTR, a novel inference-time defense that mitigates multimodal jailbreak attacks through optimizing the model's key-value (KV) caches. Rather than relying on curated safety-specific data or costly image-to-text conversion, we introduce a new formulation of the safety-relevant distributional shift induced by the visual modality. This formulation enables DTR to dynamically adjust visual token weights, minimizing the impact of adversarial visual inputs while preserving the model's general capabilities and inference efficiency. Extensive evaluation across diverse VLMs and attack benchmarks demonstrates that \sys outperforms existing defenses in both attack robustness and benign task performance, marking the first successful application of KV cache optimization for safety enhancement in multimodal foundation models. (warning: this paper contains potentially harmful content generated by VLMs.)

Jiacheng Liang, Zian Wang, Lauren Hong et al.

Various watermarking methods (``watermarkers'') have been proposed to identify LLM-generated texts; yet, due to the lack of unified evaluation platforms, many critical questions remain under-explored: i) What are the strengths/limitations of various watermarkers, especially their attack robustness? ii) How do various design choices impact their robustness? iii) How to optimally operate watermarkers in adversarial environments? To fill this gap, we systematize existing LLM watermarkers and watermark removal attacks, mapping out their design spaces. We then develop WaterPark, a unified platform that integrates 10 state-of-the-art watermarkers and 12 representative attacks. More importantly, by leveraging WaterPark, we conduct a comprehensive assessment of existing watermarkers, unveiling the impact of various design choices on their attack robustness. We further explore the best practices to operate watermarkers in adversarial environments. We believe our study sheds light on current LLM watermarking techniques while WaterPark serves as a valuable testbed to facilitate future research.

Jiacheng Liang, Songze Li, Bochuan Cao et al.

We propose OmniLytics, a blockchain-based secure data trading marketplace for machine learning applications. Utilizing OmniLytics, many distributed data owners can contribute their private data to collectively train an ML model requested by some model owners, and receive compensation for data contribution. OmniLytics enables such model training while simultaneously providing 1) model security against curious data owners; 2) data security against the curious model and data owners; 3) resilience to malicious data owners who provide faulty results to poison model training; and 4) resilience to malicious model owners who intend to evade payment. OmniLytics is implemented as a blockchain smart contract to guarantee the atomicity of payment. In OmniLytics, a model owner splits its model into the private and public parts and publishes the public part on the contract. Through the execution of the contract, the participating data owners securely aggregate their locally trained models to update the model owner's public model and receive reimbursement through the contract. We implement a working prototype of OmniLytics on Ethereum blockchain and perform extensive experiments to measure its gas cost, execution time, and model quality under various parameter combinations. For training a CNN on the MNIST dataset, the MO is able to boost its model accuracy from 62% to 83% within 500ms in blockchain processing time.This demonstrates the effectiveness of OmniLytics for practical deployment.