arXiv subjects

Yunhao Feng

Publications and source records attributed to Yunhao Feng.

18 recordsLinked to original sources

arXiv2026

MemHarmBench: Evaluating Cross-Session Prompt Injection in Memory-Persistent AI Agents

AI agents often rely on persistent memory to retain user preferences and operational rules across sessions. However, this capability introduces a new safety risk: malicious instructions from messages, documents, or Web content can be retained in memory and influence future tasks even after the original attack content is removed. Existing attack success metrics, however, fail to capture who is affected and how attacks translate into downstream harm during otherwise normal task execution. We introduce MemHarmBench, a victim-centric benchmark for evaluating cross-session prompt injection in memory-persistent AI agents. MemHarmBench considers three categories of potential victims: users, workspace/project owners, and external stakeholders. Our evaluation first exposes an agent to malicious instructions and tests whether they are retained in memory. We then remove the attack content and initiate a new session with a legitimate task while preserving the agent's memory. We jointly evaluate whether the agent follows the attacker's intent, completes the legitimate task, and exhibits observable execution anomalies. Across multiple agent systems, we find that cross-session attacks can remain behaviorally stealthy: agents often complete tasks normally and exhibit no detectable anomalies while still making critical decisions aligned with the attacker's intent. By connecting malicious memory retention to downstream decisions and affected parties, MemHarmBench provides a systematic framework for exposing safety risks that conventional attack- and task-success metrics may overlook. We publicly release MemHarmBench at https://github.com/Kirito14IT/MemHarmBench.

cs.AI↗

arXiv2026

CST: Lightweight Risk Triage with Source-Grounded Evidence for Web Discussions

Web communities rely on moderators to catch harmful messages such as insults, hate speech, and scams, but no one can read every message. Existing tools give each message a risk score, which tells moderators where to look but not what to read or how to check it. Yet meaning depends on context: the same reply, "I agree.", can support or oppose removing someone from a group. Moderators therefore need source-grounded evidence, not just scores: a few key messages kept with their original wording and replies. We present CST, a lightweight triage pipeline that first decides where to look, then what to show. It scores risk by pairing character patterns, which capture exact wording, with sentence meaning from a fixed multilingual model, so it handles several languages and runs on an ordinary CPU. It then groups related messages and picks from each group in turn, so a small reading budget covers different parts of a discussion, and links every pick back to the original thread. Across six tasks from five datasets in several languages, CST ranks harmful messages more accurately than a fully fine-tuned multilingual model and, under the same budget, finds more of them than reading by score alone. In a user study, reviewers needed at least 27% less time with CST's grouped evidence while keeping decision accuracy above 94%. CST turns risk scores into evidence that moderators can read quickly and check in context. Code and dataset reconstruction resources will be released at https://github.com/Kirito14IT/CST-MIL.

cs.CL↗

arXiv2026

Async-RBench: Benchmarking Asynchronous Task Execution in LLM Agents

Large language model (LLM) agents increasingly execute tasks asynchronously: a main agent continues working while its subagents run, and the task state may change before their results arrive. The main agent must decide which work to update or preserve and which checks to repeat, while avoiding prohibited changes. Final task success alone does not reveal whether these results were handled correctly, and natural asynchronous arrivals are hard to reproduce. We introduce Async-RBench, a benchmark for asynchronous task execution in LLM agents, with 200 executable tasks covering eight scenarios. A scenario controller regulates result delivery using predefined task-state conditions, while an Event Response Contract (ERC) specifies what each response must change, preserve, avoid, and re-verify. The Dynamic Replanning Score (DRS) weights the response process and outcome equally and is reported separately from task success. We evaluate nine agents and find DRS scores of 19.2 to 45.1, with the lowest scores on duplicate reports and subagent failures. Under our execution protocol, all nine agents obtain lower task success with asynchronous than with batched delivery (26.0% vs. 29.7% on average). For three agents, prompting the main agent to rebuild its plan after every result improves DRS by only 2.6 to 3.5 points. Together, these findings highlight the challenges of reliable asynchronous task execution and the need to evaluate how agents handle incoming results alongside end-to-end task success.

cs.AI↗

arXiv2026

AdaGuard: An Adaptive Guard Model with User-defined Policies

Guard models support the safe deployment of language model agents, but fixed risk taxonomies limit their ability to accommodate requirements that vary across applications and tasks. Under user-defined policies, detecting violations requires interpreting both the applicable rules and the agent's behavior, since identical actions can receive different judgments under different policies. To support learning this capability, we introduce AdaptiveSafety, a dataset of 10,939 training examples and 1,000 test examples covering policies with 1--100 rules. The dataset combines trajectories from multiple sources with policy and behavioral counterfactuals, pairing each example with an explanation and the complete set of violated rules. These counterfactuals expose changes that alter compliance, while structural augmentations provide supervision for consistency under rule reordering and identifier remapping. Building on this supervision, we propose SafePO, a reinforcement learning algorithm for refining violation identification while balancing explanatory reasoning and final verdicts. SafePO uses structured rewards to assess prediction correctness, retains group-relative advantages at the response level, and employs a separately trained value model to modulate token weights within explanation and verdict regions. Separate normalization controls their relative contribution to training despite differences in length. Through supervised initialization followed by SafePO, we develop AdaGuard, a family of 0.6B, 4B, and 8B guard models that assess agent trajectories under policies supplied at inference time. Our 4B model achieves binary accuracies of 89.30\% on AdaptiveSafety and 71.82\% on DynaBench. The project repository is available at https://github.com/Yunhao-Feng/AdaGuard

cs.AI↗

arXiv2026

ShadowFPT: Backdooring Federated Prompt Tuning via Shadow Triggers

Federated Prompt Tuning (FPT) adapts large vision--language models by freezing the pretrained backbone and optimizing only lightweight prompt parameters across clients. Although this design improves communication and parameter efficiency, it creates an overlooked security risk. Since the backbone is shared and frozen, a malicious client can induce backdoor behavior through the representation space while keeping its uploaded prompt updates close to benign ones. We propose \textbf{ShadowFPT}, a targeted backdoor attack that exploits this frozen-backbone attack surface. ShadowFPT first pretrains a learnable \emph{Shadow Trigger} against the frozen CLIP visual encoder, using either auxiliary public data or the malicious client's local data, so that triggered inputs are steered toward the target class in representation space. During federated prompt tuning, the malicious client adapts the trigger under the current global prompt and then optimizes its local prompt on both clean and triggered samples. Only prompt parameters are uploaded to the server, while the trigger and frozen encoders remain local. By shifting most of the attack burden from prompt updates to trigger-induced representation steering, ShadowFPT achieves targeted misclassification while preserving prompt-space stealthiness. Across multiple datasets, aggregation rules, and non-IID federated partitions, ShadowFPT increases the attack success rate from 19.81\% to 90.36\% in our main setting, while maintaining clean accuracy. It remains effective across textual, visual, and joint vision--language prompt tuning. These results identify frozen backbones as stealthy and underexplored backdoor surfaces in federated prompt tuning, suggesting that defenses based only on prompt-update anomaly detection are insufficient.

cs.AI↗

arXiv2026

AgentHazard: A Benchmark for Evaluating Harmful Behavior in Computer-Use Agents

Computer-use agents extend language models from text generation to persistent action over tools, files, and execution environments. Unlike chat systems, they maintain state across interactions and translate intermediate outputs into concrete actions. This creates a distinct safety challenge in that harmful behavior may emerge through sequences of individually plausible steps, including intermediate actions that appear locally acceptable but collectively lead to unauthorized actions. We present \textbf{AgentHazard}, a benchmark for evaluating harmful behavior in computer-use agents. AgentHazard contains \textbf{2,653} instances spanning diverse risk categories and attack strategies. Each instance pairs a harmful objective with a sequence of operational steps that are locally legitimate but jointly induce unsafe behavior. The benchmark evaluates whether agents can recognize and interrupt harm arising from accumulated context, repeated tool use, intermediate actions, and dependencies across steps. We evaluate AgentHazard on Claude Code, OpenClaw, and IFlow using mostly open or openly deployable models from the Qwen3, Kimi, GLM, and DeepSeek families. Our experimental results indicate that current systems remain highly vulnerable. In particular, when powered by Qwen3-Coder, Claude Code exhibits an attack success rate of \textbf{73.63\%}, suggesting that model alignment alone does not reliably guarantee the safety of autonomous agents.

cs.AI↗

arXiv2026

HazardAuditor: From Executable Threats to Safer Computer-Use Agents

Computer-use agents increasingly interact with browsers, terminals, file systems, and external services, introducing safety risks that emerge through runtime behavior rather than generated content alone. Existing guard models target static prompts and responses and are poorly suited to agent execution; existing executable safety platforms produce evaluation verdicts rather than the normalized supervision a guard model needs to learn across heterogeneous agent frameworks. We introduce HazardAuditor, an execution-grounded framework that closes both gaps. Its infrastructure runs heterogeneous agents (Claude Code, Codex, Hermes, and OpenClaw) in controlled environments and normalizes their interactions into a canonical event representation for cross-framework supervision. We further observe that token-level post-training objectives create a structural mismatch for generative guards, causing longer rationales to dominate gradient updates. Guard Policy Optimization (GuardPO) addresses this by converting deterministic safety outcomes into sequence-level advantages and normalizing rationale and verdict regions, making the safety decision the effective unit of optimization. Across multiple benchmarks and heterogeneous computer-use systems, HazardAuditor improves accuracy by up to 16.5 percentage points over the strongest prior guard. Code, models, and evaluation artifacts will be available at https://yunhao-feng.github.io/HazardAuditor/.

cs.AI↗

arXiv2026

Constitutional On-Policy Safe Distillation

On-policy self-distillation (OPSD) has emerged as an efficient post-training paradigm by using a teacher conditioned on privileged information to provide dense token-level supervision. Prior work has shown that OPSD can collapse in verifiable reasoning tasks, while safety alignment differs in that it is guided by high-level constitutions rather than explicit target answers. However, pilot studies reveals that safety OPSD nonetheless suffers from severe collapse, where constitutional conditioning contracts the teacher distribution toward short and overly conservative responses and Reverse KL further amplifies this contraction into reduced expressiveness. We formalize this effect as geometric leakage under safety boundaries in a non-orthogonal semantic space, where safety pressure transfers into the expressiveness dimension. Based on this analysis, we propose Constitutional On-Policy Safety Distillation (COPSD), which first calibrates the teacher through a Cross-SFT cold-start and then performs constitution-conditioned on-policy distillation. Experiments across 3 multimodal large language models (MLLMs) on 12 safety and general benchmarks show that COPSD improves both safety and helpfulness over baselines while reducing the safety tax on general reasoning.

cs.LG↗

arXiv2026

SkillTrace: Multi-Trace Provenance Auditing for LLM-Agent Skill Reuse

LLM-agent ecosystems are rapidly growing around reusable skills: mixed-modality packages of metadata, natural-language instructions, code, tools, references, and operational workflows. As skills become marketplace artifacts, auditing their reuse is no longer the same problem as ordinary code clone detection. Existing detectors target single-modality source code or whole-package similarity, yet skill reuse evidence is distributed across authored text, implementation fragments, and operational structure. As a result, they can miss reuse that preserves only one part of a skill. We present SKILLTRACE, a multi-trace provenance auditing framework for LLM-agent skill reuse. SKILLTRACE extracts three provenance traces: Expression, Implementation, and Operational. It represents the Operational Trace as a Skill Operational Graph (SOG) that captures activation, procedure, and resource-flow structure. An LLM assists only the Operational-trace extraction, once at ingestion; at audit time SKILLTRACE compares cached traces deterministically, calibrates each trace against same-function strict negatives, and reports which trace supports a reuse decision. On SKILLTRACE-BENCH, with 820 transformed reuse positives over 100 marketplace anchors and 751 negative controls, SKILLTRACE achieves AUROC 0.938 and F1 0.898. A 36,446-skill wild audit further shows that trace-attributed evidence surfaces actionable reuse review queues beyond repository-level baselines.

cs.AI↗

arXiv2026

When Experience Becomes Instruction: Trajectory Poisoning in Self-Evolving Agent Skill Systems

Self-evolving skill (SES) systems distill agent trajectories into persistent skills, allowing untrusted experience to become trusted instruction. We introduce PoisonedEvolution, a trajectory-poisoning attack on this promotion process. Our skill-visible black-box attacker can inspect a target skill and contribute bounded evidence, but cannot observe private pools or evolution logic or edit the skill bank. Artifact poisoning requires Inclusion, Evolution Attribution, and Realization. Attribution is the distinctive bottleneck: the target behavior must appear causally useful, recurrent, and generalizable before promotion. We evaluate four representative security-effect families using inert canary specifications. At 10% attacker support, across six mainstream LLM evolvers in SkillClaw, PoisonedEvolution embeds target behaviors in 546/600 trials (91.0% SER). On the structurally different Trace2Skill pipeline at the same ratio, it embeds target behaviors in 369/600 trials (61.5% SER), demonstrating transfer across evolution architectures. In a representative controlled study, three consistent attacker records suffice in a 30-record batch, whereas a single record is much weaker. Ablations identify recurring support, causal framing, and domain-aligned encoding as the main determinants of success. These findings expose evidence promotion as a security boundary for self-evolving agents.

cs.CR↗

arXiv2026

ASEval: Automated Trajectory-Level Security Testing for Autonomous Agents

As autonomous agents (e.g., OpenClaw) increasingly operate with deep system-level privileges to execute complex tasks, they introduce severe, unmitigated security risks. Existing LLM safety testing methods are largely built around prompt-level inputs and response-level judgments, while recent agent benchmarks remain limited in automation, trajectory coverage, and action-grounded judgment. In this work, we present ASEval, a novel automated security testing framework for autonomous agents. ASEval generates realistic benign multi-turn conversations, perturbs them into risk test cases through carefully designed perturbation operators, executes each case against the target agent, and applies an action-grounded oracle to judge the complete trace. We instantiate ASEval to construct A3S-Bench, a test suite of 2,254 executable multi-turn test cases spanning diverse usage scenarios, delivery patterns, and agent-level security risks. Extensive evaluation on 11 LLM-backed OpenClaw agents confirms that ASEval effectively exposes pervasive security failures. In particular, our proposed perturbation operators raise the average risk-trigger rate from 22.9% under existing injection techniques to 47.4%, as judged by an action-grounded oracle that matches human annotation at 93.2% binary F1. Our further analysis and case studies illustrate that such failures emerge from multi-turn delivery, persistent state, and tool-mediated actions that response-level evaluation overlooks.

cs.CR↗

arXiv2026

Safety Testing LLM Agents at Scale: From Risk Discovery to Evidence-Grounded Verification

LLM agents increasingly perform autonomous actions through external tools, leading to complex and evolving safety risks. However, existing safety testing targets expert-designed safety violations, and the corresponding outcomes are evaluated by hard-coded rules, making them costly to extend as agents evolve. To this end, we present Vera, an end-to-end automated safety testing framework that instantiates software engineering testing principles for non-deterministic agents through a three-stage, self-reinforcing pipeline. First, a literature-driven exploration continuously discovers and structures emerging risks into taxonomies of safety risks, attack methods, and tool execution environments. Second, combinatorial composition across taxonomy dimensions produces executable safety cases, each specifying a concrete safety goal, a programmatically constructed initial state, and a deterministic verification predicate grounded in observable artifacts. Third, adaptive execution runs heterogeneous agents in isolated sandboxes where a control agent steers multi-turn interaction based on runtime observations, while evidence-grounded verifiers judge outcomes from environment state and tool-call evidence rather than model self-report. We evaluate Vera on four production agent frameworks (OpenClaw, Hermes, Codex, Claude Code), revealing substantial safety weaknesses, with average attack success rates reaching 93.9\% under multi-channel attacks; we also release Vera-Bench, comprising 1600 executable safety cases spanning 124 risk categories across three execution settings. These results indicate that modular, executable testing infrastructure is essential for rigorous and maintainable safety evaluation of rapidly evolving agentic systems at scale. The code is publicly available at https://github.com/Yunhao-Feng/Vera.

cs.AI↗

arXiv2026

Safety in Self-Evolving LLM Agent Systems: Threats, Amplification, and Case Studies

Self-evolving LLM agent systems, which autonomously update their model parameters, memory, tools, and architectures, introduce a qualitatively new threat landscape in which adversarial influences become permanently encoded, self-amplify across generations, and propagate through populations without sustained attacker access. We present a systematic security and privacy analysis organized around the Module-Lifecycle Attack Surface (MLAS) matrix, which decomposes the attack surface into five functional modules (Brain, Cognitive Resource, Execution, Self-Design, Collective) $\times$ five lifecycle stages (Bootstrap, Propose, Evaluate, Commit, Serve). Analysis of the resulting 25 cells reveals that 17 face critical threats for which no effective partial mitigation. We identify seven cross-cutting amplification effects that interact synergistically and cannot be addressed by securing individual modules in isolation. Comparative case studies of two open-source frameworks demonstrate that evolution-native design activates $3.5\times$ more attack surface cells and achieves a 100% attack persistence rate (40/40 payloads across all CIA+Privacy categories), while co-located security scanners block only 2.5% of attacks. Our findings establish that self-evolution converts every known attack category from session-bounded to lineage-persistent, gives rise to entirely new attack classes, and renders static defenses structurally inadequate, motivating evolution-aware security frameworks and formal verification for self-modifying systems.

cs.CR↗

arXiv2026

BraveGuard: From Open-World Threats to Safer Computer-Use Agents

Computer-use agents extend language models from text generation to sustained interaction with files, terminals, browsers, and external tools. This shift creates safety risks that are difficult to detect from isolated prompts or final responses, because harm often emerges only through multi-step execution traces whose individual actions appear locally benign. We introduce BraveGuard, a self-evolving defense framework for training guard models from open-world threat signals and realistic agent trajectories. BraveGuard mines recent research sources to identify emerging risks and attack patterns, instantiates them as executable computer-use tasks, collects agent rollouts, and derives trajectory-level supervision for guard model training. As new threats and validation failures appear, the pipeline can be repeated, yielding an adaptive defense loop rather than a static, benchmark-driven training process. We instantiate BraveGuard by training multiple guard backbones, including Qwen3-Guard and Llama-Guard variants, and evaluate the resulting guards on trajectory-level agent-safety benchmarks. BraveGuard consistently improves safety detection across computer-use trajectories. On AgentHazard, it substantially improves detection accuracy over off-the-shelf guard models, with accuracy increasing from 38.79% to 82.38% under the averaged guard-model setting. These results show that guard supervision grounded in open-world threat discovery and realistic agent execution can improve safety monitoring beyond fixed taxonomies and synthetic prompt-level data. BraveGuard offers a scalable path toward adaptive defenses for computer-use agents facing evolving real-world risks.

cs.CR↗

arXiv2026

SkillTrojan: Backdoor Attacks on Skill-Based Agent Systems

Skill-based agent systems tackle complex tasks by composing reusable skills, improving modularity and scalability while introducing a largely unexamined security attack surface. We propose SkillTrojan, a backdoor attack that targets skill implementations rather than model parameters or training data. SkillTrojan embeds malicious logic inside otherwise plausible skills and leverages standard skill composition to reconstruct and execute an attacker-specified payload. The attack partitions an encrypted payload across multiple benign-looking skill invocations and activates only under a predefined trigger. SkillTrojan also supports automated synthesis of backdoored skills from arbitrary skill templates, enabling scalable propagation across skill-based agent ecosystems. To enable systematic evaluation, we release a dataset of 3,000+ curated backdoored skills spanning diverse skill patterns and trigger-payload configurations. We instantiate SkillTrojan in a representative code-based agent setting and evaluate both clean-task utility and attack success rate. Our results show that skill-level backdoors can be highly effective with minimal degradation of benign behavior, exposing a critical blind spot in current skill-based agent architectures and motivating defenses that explicitly reason about skill composition and execution. Concretely, on EHR SQL, SkillTrojan attains up to 97.2% ASR while maintaining 89.3% clean ACC on GPT-5.2-1211-Global.

cs.CR↗

arXiv2026

Position: AI Safety Requires Effective Controllability

AI safety is still largely framed as alignment: training models to follow human preferences, safety policies, and normative constraints. That framing has improved the behavior of modern language models, but aligned behavior does not by itself guarantee that a deployed agent can be stopped, overridden, or constrained once it operates in open-ended, interactive, and tool-using environments. A system may be safe in expectation and still fail to yield to explicit runtime authority under conflicting instructions, long-horizon execution, adversarial inputs, or risky tool use. This position paper argues that AI safety therefore requires controllability as a first-class objective. We define \emph{controllability} as the ability of an AI system to remain reliably interruptible, overridable, redirectable, and constrainable by explicit control signals at runtime while preserving ordinary utility when such signals are absent. To study this gap, we introduce \controlbench{}, a benchmark for evaluating controllability failures in high-risk agentic scenarios. Experiments with OpenClaw-based agents show that current alignment and guardrail mechanisms reduce risk, but often fail to provide persistent, authoritative, and enforceable runtime control. We therefore propose a control-centric architectural framework that highlights explicit control planes, runtime intervention pathways, persistent control states, and auditable decision interfaces as key design principles for future controllable AI systems.

cs.AI↗

arXiv2026

BackdoorAgent: A Unified Framework for Backdoor Attacks on LLM-based Agents

Large language model (LLM) agents execute tasks through multi-step workflows that combine planning, memory, and tool use. While this design enables autonomy, it also expands the attack surface for backdoor threats. Backdoor triggers injected into specific stages of an agent workflow can persist through multiple intermediate states and adversely influence downstream outputs. However, existing studies remain fragmented and typically analyze individual attack vectors in isolation, leaving the cross-stage interaction and propagation of backdoor triggers poorly understood from an agent-centric perspective. To fill this gap, we propose \textbf{BackdoorAgent}, a modular and stage-aware framework that provides a unified, agent-centric view of backdoor threats in LLM agents. BackdoorAgent structures the attack surface into three functional stages of agentic workflows, including \textbf{planning attacks}, \textbf{memory attacks}, and \textbf{tool-use attacks}, and instruments agent execution to enable systematic analysis of trigger activation and propagation across different stages. Building on this framework, we construct a standardized benchmark spanning four representative agent applications: \textbf{Agent QA}, \textbf{Agent Code}, \textbf{Agent Web}, and \textbf{Agent Drive}, covering both language-only and multimodal settings. Our empirical analysis shows that \textit{triggers implanted at a single stage can persist across multiple steps and propagate through intermediate states.} For instance, when using a GPT-based backbone, we observe trigger persistence in 43.58\% of planning attacks, 77.97\% of memory attacks, and 60.28\% of tool-stage attacks, highlighting the vulnerabilities of the agentic workflow itself to backdoor threats. To facilitate reproducibility and future research, our code and benchmark are publicly available at GitHub.

cs.AI↗

arXiv2025

PaperAsk: A Benchmark for Reliability Evaluation of LLMs in Paper Search and Reading

Large Language Models (LLMs) increasingly serve as research assistants, yet their reliability in scholarly tasks remains under-evaluated. In this work, we introduce PaperAsk, a benchmark that systematically evaluates LLMs across four key research tasks: citation retrieval, content extraction, paper discovery, and claim verification. We evaluate GPT-4o, GPT-5, and Gemini-2.5-Flash under realistic usage conditions-via web interfaces where search operations are opaque to the user. Through controlled experiments, we find consistent reliability failures: citation retrieval fails in 48-98% of multi-reference queries, section-specific content extraction fails in 72-91% of cases, and topical paper discovery yields F1 scores below 0.32, missing over 60% of relevant literature. Further human analysis attributes these failures to the uncontrolled expansion of retrieved context and the tendency of LLMs to prioritize semantically relevant text over task instructions. Across basic tasks, the LLMs display distinct failure behaviors: ChatGPT often withholds responses rather than risk errors, whereas Gemini produces fluent but fabricated answers. To address these issues, we develop lightweight reliability classifiers trained on PaperAsk data to identify unreliable outputs. PaperAsk provides a reproducible and diagnostic framework for advancing the reliability evaluation of LLM-based scholarly assistance systems.

cs.IR↗