arXiv subjects

Zesen Liu

Publications and source records attributed to Zesen Liu.

10 recordsLinked to original sources

arXiv2026

Safe to Check, Unsafe to Use: Relinking at the Compression Boundary of LLM Agents

Summarization-based prompt compression is increasingly used by LLM agents to shorten long, distributed contexts, but it shifts the security boundary: filters inspect the pre-compression prompt while the backend acts on a newly generated compressed context. We identify relinking, a compression-boundary vulnerability where the compressor behaves as a confused deputy, summarizing distributed, locally benign fragments into a complete malicious instruction. Unlike prompt injection, relinking need not place an explicitly malicious payload in the source context. We show that relinking arises from summarization itself: attention makes separated fragments jointly available, pre-training makes compatible fragments plausible to connect, and post-training favors compact backend-actionable summaries. We formalize the attacker-induced form as adversarial relinking and present Relink, an automated DSL-based tool that splits malicious payloads into benign fragments while keeping the complete payload absent before compression. Across four long-context agent benchmarks, Relink achieves 86.9% Relink Rate and Backend Action Rate versus 17.0% for clean-split controls. Existing defenses fail to reliably capture adversarial relinking; our KBRA defense reduces residual Backend Action Rate to 0.0%.

cs.CR↗

arXiv2026

Rewriting the Response Path: Silent Tampering and Provider-Signed Defense in BYOK LLM Agents

LLM agents convert model outputs into consequential actions, including communications, code changes, and financial transactions. Developers often trust evidence such as test results and execution logs. We identify a response path integrity gap in Bring Your Own Key configurations used by roughly 88 percent of mainstream agents. Because traffic passes through a user-authorized relay, the relay can modify plaintext LLM responses after alignment but before execution without breaking encryption. A minimal attack rewrites one execution bearing field and regenerates the remaining response using the user key while preserving the model style. Experiments reveal false green verification, where malicious code modifications pass public tests while silently defeating security checks. On APPS, 99.7 percent of publicly passing solutions retained downgraded behavior without developer-visible warnings. Tests on SWE bench, AgentDojo, and ASB across five frontier models show that single-field rewriting can redirect agents while preserving apparent task completion. We propose sign-c, a server-side scheme that authenticates execution bearing fields and outgoing queries. A local shim verifies them before action, while encryption protects confidentiality. The defense rejected all tampered responses with zero false rejections and only 0.0167 percent latency overhead.

cs.CR↗

arXiv2026

From Similarity to Vulnerability: Key Collision Attack on LLM Semantic Caching

Semantic caching has emerged as a pivotal technique for scaling LLM applications, widely adopted by major providers including AWS and Microsoft. By utilizing semantic embedding vectors as cache keys, this mechanism effectively minimizes latency and redundant computation for semantically similar queries. In this work, we conceptualize semantic cache keys as a form of fuzzy hashes. We demonstrate that the locality required to maximize cache hit rates fundamentally conflicts with the cryptographic avalanche effect necessary for collision resistance. Our conceptual analysis formalizes this inherent trade-off between performance (locality) and security (collision resilience), revealing that semantic caching is naturally vulnerable to key collision attacks. While prior research has focused on side-channel and privacy risks, we present the first systematic study of integrity risks arising from cache collisions. We introduce CacheAttack, an automated framework for launching black-box collision attacks. We evaluate CacheAttack in security-critical tasks and agentic workflows. It achieves a hit rate of 86\% in LLM response hijacking and can induce malicious behaviors in LLM agent, while preserving strong transferability across different embedding models. A case study on a financial agent further illustrates the real-world impact of these vulnerabilities. Finally, we discuss mitigation strategies.

cs.CR↗

arXiv2025

When Compression Becomes an Attack Surface: Black-Box Attacks on Prompt-Compressed LLM Agents

Prompt compression is increasingly deployed in LLM agents to reduce latency and cost, but it also determines what the backend LLM ultimately sees. We show that, when trusted and untrusted inputs are compressed under a shared budget, this lossy transformation creates a new attack surface: by perturbing only untrusted inputs before compression, an adversary can cause the compressor to discard task-critical evidence or safety guardrails before inference. Unlike prompt injection, jailbreaks, or RAG poisoning, the attack target is the compressor rather than the backend LLM; the perturbation need not encode a meaningful instruction or survive compression. We formalize this vulnerability as adversarial information loss (AIL), the excess downstream distortion caused by adversarially steering a lossy compressor beyond benign compression alone. To exploit AIL, we present COMA, a transfer-based black-box attack that optimizes pre-compression perturbations using attacker-side surrogate compressors and backend LLMs. Across three tasks and six compressors, COMA achieves 0.71 average ASR, versus 0.21 for the strongest baseline, and transfers to two real-world agent case studies.

cs.CR↗

arXiv2025

QueryIPI: Query-agnostic Indirect Prompt Injection on Coding Agents

Modern coding agents integrated into IDEs orchestrate powerful tools and high-privilege system access, creating a high-stakes attack surface. Prior work on Indirect Prompt Injection (IPI) is mainly query-specific, requiring particular user queries as triggers and leading to poor generalizability. We propose query-agnostic IPI, a new attack paradigm that reliably executes malicious payloads under arbitrary user queries. Our key insight is that malicious payloads should leverage the invariant prompt context (i.e., system prompt and tool descriptions) rather than variant user queries. We present QueryIPI, an automated framework that uses tool descriptions as optimizable payloads and refines them via iterative, prompt-based blackbox optimization. QueryIPI leverages system invariants for initial seed generation aligned with agent conventions, and iterative reflection to resolve instruction-following failures and safety refusals. Experiments on five simulated agents show that QueryIPI achieves up to 87% success rate, outperforming the best baseline (50%). Crucially, generated malicious descriptions transfer to real-world coding agents, highlighting a practical security risk.

cs.CR↗

arXiv2025

Red-Teaming Coding Agents from a Tool-Invocation Perspective: An Empirical Security Assessment

Coding agents powered by large language models are becoming central modules of modern IDEs, helping users perform complex tasks by invoking tools. While powerful, tool invocation opens a substantial attack surface. Prior work has demonstrated attacks against general-purpose and domain-specific agents, but none have focused on the security risks of tool invocation in coding agents. To fill this gap, we conduct the first systematic red-teaming of six popular real-world coding agents: Cursor, Claude Code, Copilot, Windsurf, Cline, and Trae. Our red-teaming proceeds in two phases. In Phase 1, we perform prompt leakage reconnaissance to recover system prompts. We discover a general vulnerability, ToolLeak, which allows malicious prompt exfiltration through benign argument retrieval during tool invocation. In Phase 2, we hijack the agent's tool-invocation behavior using a novel two-channel prompt injection in the tool description and return values, achieving remote code execution (RCE). We adaptively construct payloads using security information leaked in Phase 1. In emulation across five backends, our method outperforms baselines on Claude-Sonnet-4, Claude-Sonnet-4.5, Grok-4, and GPT-5. On real agents, our approach succeeds on 19 of 25 agent-LLM pairs, achieving leakage on every agent using Claude and Grok backends. For tool-invocation hijacking, we obtain RCE on every tested agent-LLM pair, with our two-channel method delivering the highest success rate. We provide case studies on Cursor and Claude Code, analyze security guardrails of external and built-in tools, and conclude with practical defense recommendations.

cs.CR↗

arXiv2024

On Evaluating The Performance of Watermarked Machine-Generated Texts Under Adversarial Attacks

Large Language Models (LLMs) excel in various applications, including text generation and complex tasks. However, the misuse of LLMs raises concerns about the authenticity and ethical implications of the content they produce, such as deepfake news, academic fraud, and copyright infringement. Watermarking techniques, which embed identifiable markers in machine-generated text, offer a promising solution to these issues by allowing for content verification and origin tracing. Unfortunately, the robustness of current LLM watermarking schemes under potential watermark removal attacks has not been comprehensively explored. In this paper, to fill this gap, we first systematically comb the mainstream watermarking schemes and removal attacks on machine-generated texts, and then we categorize them into pre-text (before text generation) and post-text (after text generation) classes so that we can conduct diversified analyses. In our experiments, we evaluate eight watermarks (five pre-text, three post-text) and twelve attacks (two pre-text, ten post-text) across 87 scenarios. Evaluation results indicate that (1) KGW and Exponential watermarks offer high text quality and watermark retention but remain vulnerable to most attacks; (2) Post-text attacks are found to be more efficient and practical than pre-text attacks; (3) Pre-text watermarks are generally more imperceptible, as they do not alter text fluency, unlike post-text watermarks; (4) Additionally, combined attack methods can significantly increase effectiveness, highlighting the need for more robust watermarking solutions. Our study underscores the vulnerabilities of current techniques and the necessity for developing more resilient schemes.

cs.CR↗

arXiv2024

Have You Merged My Model? On The Robustness of Large Language Model IP Protection Methods Against Model Merging

Model merging is a promising lightweight model empowerment technique that does not rely on expensive computing devices (e.g., GPUs) or require the collection of specific training data. Instead, it involves editing different upstream model parameters to absorb their downstream task capabilities. However, uncertified model merging can infringe upon the Intellectual Property (IP) rights of the original upstream models. In this paper, we conduct the first study on the robustness of IP protection methods under model merging scenarios. Specifically, we investigate two state-of-the-art IP protection techniques: Quantization Watermarking and Instructional Fingerprint, along with various advanced model merging technologies, such as Task Arithmetic, TIES-MERGING, and so on. Experimental results indicate that current Large Language Model (LLM) watermarking techniques cannot survive in the merged models, whereas model fingerprinting techniques can. Our research aims to highlight that model merging should be an indispensable consideration in the robustness assessment of model IP protection techniques, thereby promoting the healthy development of the open-source LLM community. Our code is available at https://github.com/ThuCCSLab/MergeGuard.

cs.CR↗

arXiv2023

Fast and Adaptive Multi-agent Planning under Collaborative Temporal Logic Tasks via Poset Products

Efficient coordination and planning is essential for large-scale multi-agent systems that collaborate in a shared dynamic environment. Heuristic search methods or learning-based approaches often lack the guarantee on correctness and performance. Moreover, when the collaborative tasks contain both spatial and temporal requirements, e.g., as Linear Temporal Logic (LTL) formulas, formal methods provide a verifiable framework for task planning. However, since the planning complexity grows exponentially with the number of agents and the length of the task formula, existing studies are mostly limited to small artificial cases. To address this issue, a new planning paradigm is proposed in this work for system-wide temporal task formulas that are released online and continually. It avoids two common bottlenecks in the traditional methods, i.e., (i) the direct translation of the complete task formula to the associated B\"uchi automaton; and (ii) the synchronized product between the B\"uchi automaton and the transition models of all agents. Instead, an adaptive planning algorithm is proposed that computes the product of relaxed partially-ordered sets (R-posets) on-the-fly, and assigns these subtasks to the agents subject to the ordering constraints. It is shown that the first valid plan can be derived with a polynomial time and memory complexity w.r.t. the system size and the formula length. Our method can take into account task formulas with a length of more than 400 and a fleet with more than $400$ agents, while most existing methods fail at the formula length of 25 within a reasonable duration. The proposed method is validated on large fleets of service robots in both simulation and hardware experiments.

cs.RO↗

arXiv2022

Time Minimization and Online Synchronization for Multi-agent Systems under Collaborative Temporal Tasks

Multi-agent systems can be extremely efficient when solving a team-wide task in a concurrent manner. However, without proper synchronization, the correctness of the combined behavior is hard to guarantee, such as to follow a specific ordering of sub-tasks or to perform a simultaneous collaboration. This work addresses the minimum-time task planning problem for multi-agent systems under complex global tasks stated as Linear Temporal Logic (LTL) formulas. These tasks include the temporal and spatial requirements on both independent local actions and direct sub-team collaborations. The proposed solution is an anytime algorithm that combines the partial-ordering analysis of the underlying task automaton for task decomposition, and the branch and bound (BnB) search method for task assignment. Analyses of its soundness, completeness and optimality as the minimal completion time are provided. It is also shown that a feasible and near-optimal solution is quickly reached while the search continues within the time budget. Furthermore, to handle fluctuations in task duration and agent failures during online execution, an adaptation algorithm is proposed to synchronize execution status and re-assign unfinished subtasks dynamically to maintain correctness and optimality. Both algorithms are validated rigorously over large-scale systems via numerical simulations and hardware experiments, against several strong baselines.

cs.RO↗