Searcharxiv⌕ Search

arXiv subjects

Xuebing Zhou

Publications and source records attributed to Xuebing Zhou.

13 recordsLinked to original sources

Evolution Under Fire: Evolving Agent Safety Harnesses with Adaptive Injection Stressors

Tool-using language-model agents must act on information from untrusted emails, webpages, and documents. This creates a central security challenge: malicious instructions and legitimate task directives can arrive through the same interaction channel. Existing defenses are largely hand-designed and static, making further improvements dependent on repeated manual redesign. We introduce Evolution Under Fire, a staged framework that couples safety-harness evolution with adaptive stressor evolution. Utility-constrained, Pareto-scheduled tree search evolves complete executable safety harnesses from execution evidence while keeping the task model, benchmark, tools, and evaluator fixed. Adaptive stressor evolution then discovers utility-preserving attacks that guide a subsequent stage of harness evolution. On AgentDyn with GPT-4o-mini, an evolved harness reduces attack success from 50.4% to 3.0%, improves clean and attacked utility over the initial harness, and offers a stronger security-utility trade-off than reproduced defenses. On Agent-SafetyBench with Qwen2.5-7B-Instruct, harness evolution reduces attack success from 82.6% to 62.0% while retaining helpfulness and runtime validity. Evolved harnesses also transfer without further evolution to Qwen3.5-35B-A3B, reducing attack success from 32.1% to 0.9% while improving attacked-task utility.

cs.CR↗

EvoSkill-S1: Evolving Reusable Skills with Genetic-Inspired Operators

Reusable skills allow language-model agents to accumulate and reuse problem-solving experience. However, most existing methods improve a skill bank only indirectly through rollout-driven accumulation and heuristic reflection, often resulting in shallow and weakly controlled updates. We present EvoSkill-S1, an RL-driven framework that instead treats reusable skills themselves as the direct objects of evolution. Given hard-negative failures, EvoSkill-S1 selects parent skills and transforms them into candidate child skills through five genetic-inspired operators: revision, cross-revision, mutation, dual-reflection, and crossover. Downstream task rewards train the skill evolver to generate more effective child skills, while evaluator-guided write-back selectively retains beneficial updates. Complementary Skill Bank Fusion consolidates operator-specific banks to exploit their distinct recovery patterns, while a Procedural Skill Interface supports transfer to unseen problems. Across all reported math and code benchmarks, EvoSkill-S1 improves the corresponding initialized skill banks under both Qwen and GPT-family solvers, with relative gains of up to 12.39%. Compared with the strongest external baseline without direct skill evolution, the largest relative improvement reaches 20.39%.

cs.AI↗

PrivacyEvalClaw: A Tool-Augmented Deliberative Agent for Auditing Privacy Leakage in LLM Jailbreak Responses

Reliable evaluation is a critical bottleneck in privacy red-teaming of large language models (LLMs): determining whether a jailbreak response actually reveals personal information can be difficult when disclosures are indirect, contextual, partial, or obfuscated. Existing evaluation approaches increasingly rely on guard models and LLM-as-Judges, yet these evaluators may interpret the same response differently. To better characterize this evaluation setting, we conduct a systematic study of privacy evaluators and find substantial disagreement, particularly in ambiguous cases that require additional evidence or targeted analysis to resolve. Motivated by these findings, we introduce \textsc{PrivacyEvalClaw}, a training-free deliberative agent for evaluating privacy leakage in LLM jailbreak responses. Rather than relying on isolated single-pass judgments, PrivacyEvalClaw treats evaluators and analysis modules as tools, adaptively gathers evidence, investigates uncertain disclosures, and reconciles conflicting signals before producing a final judgment. Our primary experiments use pairwise LLM-as-Judge evaluation of complete trajectories, showing that PrivacyEvalClaw achieves favorable win rates against competing baselines across outcome, process, and overall quality, while producing explicit reasoning traces that make the evaluation process more transparent and auditable.

cs.CR↗

A Safe Action Is Not Enough: Feasible-Future Decoding for Vision-Language-Action Policies

A safe action is not necessarily a viable one. A frozen vision-language-action (VLA) policy can favor a locally admissible move that leaves no policy-supported route to safe task completion. We call this the feasibility-likelihood gap: likelihood ranks the next move, while feasibility depends on the futures it leaves open. To bring those futures into the decision, we derive the exact next-block marginal of the history-conditioned policy-environment trajectory law restricted to safe task completion. The derivation reveals a candidate-dependent feasible-future mass: its support records whether safe completion remains possible under the frozen continuation process, while its magnitude measures how much weighted safe-completion mass remains. Since exact evaluation is impractical online, we develop a selective finite-candidate approximation and establish conditions for recovering the best retained viable candidate. Our alarm-triggered, training-free reranker VICS-G lowers mean cumulative safety cost by 1.9%-57.5% across six Safety-CHORES settings while remaining within 2.5 percentage points of policy sampling in success and 0.82 steps in mean episode length. Our approach offers a promising and practical path toward safer task completion, grounded in an exact policy-relative target yet requiring neither policy retraining nor online rollouts.

cs.AI↗

MemWM: Memory-Augmented Text-Based World Model

World models are increasingly used to support planning in agents by predicting how environment states evolve in response to agent actions. Yet fluent next-state predictions can still omit task-critical facts, corrupt product attributes, or apply incorrect transition rules. To address such systematic prediction errors, we introduce MemWM, a memory-augmented text-based world model. MemWM uses world memory, a curated memory bank of transition rules, state caches, and hard-to-predict facts, to condition next-state imagination. We evaluate factual state preservation with Structured State Fidelity (SSF), which scores predicted states through benchmark-specific facts and fields. Compared with SFT, memory-augmented training improves SSF by up to 206.3%. In the full planning setting, we keep the policy model frozen and provide policy-side world skill: retrieved task-level skills and step-wise corrective guidance for action selection. Across ALFWorld, WebShop, and ScienceWorld, memory-augmented agents improve downstream success over an SFT-trained world-model agent, with up to a 65.4% relative gain. Sensitivity analyses further show that retrieved memory improves task success and efficiency under different memory and action-budget settings.

cs.AI↗

A LINDDUN-based Privacy Threat Modeling Framework for GenAI

As generative AI (GenAI) systems become increasingly prevalent across various technological stacks, the question of how such systems handle sensitive and personal data flows becomes increasingly important. Specifically, both the ability to harness and process large swaths of information as well as their stochastic nature raise key concerns related to both security and privacy. Unfortunately, while some of the traditional security threat modeling can effectively identify certain violations, privacy-related issues are often overlooked. To respond to these challenges, we introduce a novel domain-specific privacy threat modeling framework to support the privacy threat analysis of GenAI-based applications. This framework is constructed through a two-pronged approach: (1) a systematic review of the emerging literature on GenAI privacy threats, and (2) a case-driven application to a representative Chatbot system. These efforts yield a foundational GenAI privacy threat modeling framework built on LINDDUN. The new framework affects three out of the seven privacy threat types of LINDDUN and introduces 100 new GenAI examples to the knowledge base. Its effectiveness is validated on an AI Agent system, which demonstrates that a comprehensive privacy analysis can be supported by the new framework.

cs.CR↗

PII-Compass: Guiding LLM training data extraction prompts towards the target PII via grounding

The latest and most impactful advances in large models stem from their increased size. Unfortunately, this translates into an improved memorization capacity, raising data privacy concerns. Specifically, it has been shown that models can output personal identifiable information (PII) contained in their training data. However, reported PIII extraction performance varies widely, and there is no consensus on the optimal methodology to evaluate this risk, resulting in underestimating realistic adversaries. In this work, we empirically demonstrate that it is possible to improve the extractability of PII by over ten-fold by grounding the prefix of the manually constructed extraction prompt with in-domain data. Our approach, PII-Compass, achieves phone number extraction rates of 0.92%, 3.9%, and 6.86% with 1, 128, and 2308 queries, respectively, i.e., the phone number of 1 person in 15 is extractable.

cs.CR↗

PII Jailbreaking in LLMs via Activation Steering Reveals Personal Information Leakage

This paper investigates privacy jailbreaking in LLMs via steering, focusing on whether manipulating activations can bypass LLM alignment and alter response behaviors to privacy related queries (e.g., a certain public figure's sexual orientation). We begin by identifying attention heads predictive of refusal behavior for private attributes (e.g., sexual orientation) using lightweight linear probes trained with privacy evaluator labels. Next, we steer the activations of a small subset of these attention heads guided by the trained probes to induce the model to generate non-refusal responses. Our experiments show that these steered responses often disclose sensitive attribute details, along with other private information about data subjects such as life events, relationships, and personal histories that the models would typically refuse to produce. Evaluations across four LLMs reveal jailbreaking disclosure rates of at least 95%, with more than 50% on average of these responses revealing true personal information. Our controlled study demonstrates that private information memorized in LLMs can be extracted through targeted manipulation of internal activations.

cs.CR↗

PII-Scope: A Comprehensive Study on Training Data PII Extraction Attacks in LLMs

In this work, we introduce PII-Scope, a comprehensive benchmark designed to evaluate state-of-the-art methodologies for PII extraction attacks targeting LLMs across diverse threat settings. Our study provides a deeper understanding of these attacks by uncovering several hyperparameters (e.g., demonstration selection) crucial to their effectiveness. Building on this understanding, we extend our study to more realistic attack scenarios, exploring PII attacks that employ advanced adversarial strategies, including repeated and diverse querying, and leveraging iterative learning for continual PII extraction. Through extensive experimentation, our results reveal a notable underestimation of PII leakage in existing single-query attacks. In fact, we show that with sophisticated adversarial capabilities and a limited query budget, PII extraction rates can increase by up to fivefold when targeting the pretrained model. Moreover, we evaluate PII leakage on finetuned models, showing that they are more vulnerable to leakage than pretrained models. Overall, our work establishes a rigorous empirical benchmark for PII extraction attacks in realistic threat scenarios and provides a strong foundation for developing effective mitigation strategies.

cs.CL↗

PrivacyScalpel: Enhancing LLM Privacy via Interpretable Feature Intervention with Sparse Autoencoders

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language processing but also pose significant privacy risks by memorizing and leaking Personally Identifiable Information (PII). Existing mitigation strategies, such as differential privacy and neuron-level interventions, often degrade model utility or fail to effectively prevent leakage. To address this challenge, we introduce PrivacyScalpel, a novel privacy-preserving framework that leverages LLM interpretability techniques to identify and mitigate PII leakage while maintaining performance. PrivacyScalpel comprises three key steps: (1) Feature Probing, which identifies layers in the model that encode PII-rich representations, (2) Sparse Autoencoding, where a k-Sparse Autoencoder (k-SAE) disentangles and isolates privacy-sensitive features, and (3) Feature-Level Interventions, which employ targeted ablation and vector steering to suppress PII leakage. Our empirical evaluation on Gemma2-2b and Llama2-7b, fine-tuned on the Enron dataset, shows that PrivacyScalpel significantly reduces email leakage from 5.15\% to as low as 0.0\%, while maintaining over 99.4\% of the original model's utility. Notably, our method outperforms neuron-level interventions in privacy-utility trade-offs, demonstrating that acting on sparse, monosemantic features is more effective than manipulating polysemantic neurons. Beyond improving LLM privacy, our approach offers insights into the mechanisms underlying PII memorization, contributing to the broader field of model interpretability and secure AI deployment.

cs.LG↗

Privacy Checklist: Privacy Violation Detection Grounding on Contextual Integrity Theory

Privacy research has attracted wide attention as individuals worry that their private data can be easily leaked during interactions with smart devices, social platforms, and AI applications. Computer science researchers, on the other hand, commonly study privacy issues through privacy attacks and defenses on segmented fields. Privacy research is conducted on various sub-fields, including Computer Vision (CV), Natural Language Processing (NLP), and Computer Networks. Within each field, privacy has its own formulation. Though pioneering works on attacks and defenses reveal sensitive privacy issues, they are narrowly trapped and cannot fully cover people's actual privacy concerns. Consequently, the research on general and human-centric privacy research remains rather unexplored. In this paper, we formulate the privacy issue as a reasoning problem rather than simple pattern matching. We ground on the Contextual Integrity (CI) theory which posits that people's perceptions of privacy are highly correlated with the corresponding social context. Based on such an assumption, we develop the first comprehensive checklist that covers social identities, private attributes, and existing privacy regulations. Unlike prior works on CI that either cover limited expert annotated norms or model incomplete social context, our proposed privacy checklist uses the whole Health Insurance Portability and Accountability Act of 1996 (HIPAA) as an example, to show that we can resort to large language models (LLMs) to completely cover the HIPAA's regulations. Additionally, our checklist also gathers expert annotations across multiple ontologies to determine private information including but not limited to personally identifiable information (PII). We use our preliminary results on the HIPAA to shed light on future context-centric privacy research to cover more privacy regulations, social norms and standards.

cs.CL↗

IncogniText: Privacy-enhancing Conditional Text Anonymization via LLM-based Private Attribute Randomization

In this work, we address the problem of text anonymization where the goal is to prevent adversaries from correctly inferring private attributes of the author, while keeping the text utility, i.e., meaning and semantics. We propose IncogniText, a technique that anonymizes the text to mislead a potential adversary into predicting a wrong private attribute value. Our empirical evaluation shows a reduction of private attribute leakage by more than 90% across 8 different private attributes. Finally, we demonstrate the maturity of IncogniText for real-world applications by distilling its anonymization capability into a set of LoRA parameters associated with an on-device model. Our results show the possibility of reducing privacy leakage by more than half with limited impact on utility.

cs.CR↗

ObfuscaTune: Obfuscated Offsite Fine-tuning and Inference of Proprietary LLMs on Private Datasets

This work addresses the timely yet underexplored problem of performing inference and finetuning of a proprietary LLM owned by a model provider entity on the confidential/private data of another data owner entity, in a way that ensures the confidentiality of both the model and the data. Hereby, the finetuning is conducted offsite, i.e., on the computation infrastructure of a third-party cloud provider. We tackle this problem by proposing ObfuscaTune, a novel, efficient and fully utility-preserving approach that combines a simple yet effective obfuscation technique with an efficient usage of confidential computing (only 5% of the model parameters are placed on TEE). We empirically demonstrate the effectiveness of ObfuscaTune by validating it on GPT-2 models with different sizes on four NLP benchmark datasets. Finally, we compare to a naïve version of our approach to highlight the necessity of using random matrices with low condition numbers in our approach to reduce errors induced by the obfuscation.

cs.CR↗