Sanitizing

Antimicrobial peptides (AMPs) often act against multiple pathogen classes, making multi-label activity prediction a more realistic screening target than binary antimicrobial classification. The ESCAPE benchmark formalizes this setting, but leading approaches typically rely on multimodal, structure-conditioned deep models that are costly to train and tune. We show that a simple, sequence-only pipeline can match and surpass these methods by combining 330 interpretable sequence descriptors with TabPFN, a tabular foundation model that performs in-context prediction in a single forward pass without gradient-based training or hyperparameter search. On ESCAPE (82,359 peptides; five labels), a label-powerset TabPFN model achieves mAP-5 = 77.8%, improving on the previously best reported 72.1%. A probabilistic classifier chain is the first method to match or exceed the best published average precision on each of the five labels simultaneously. The gains persist under the prior state-of-the-art single-fold training protocol, indicating they are not a training-set-size artefact, and are largest for remote homologues (+11.2 points below 30% sequence identity). Ablations further show that predicted structure is unnecessary at inference and that performance is not driven by any single descriptor family: ten global physicochemical scalars recover 91% of full-feature performance. Finally, explicitly modelling label dependence yields targeted benefits for scarce activities and supports ranking which activity to assay next from partial positive evidence.

With the rapid release of open-weight large foundation models, safety threats are shifting from black-box jailbreaks to neuron-level white-box attacks that directly identify and manipulate safety-related neurons. Existing alignment methods often investigate the safety behavior on a small number of neurons, creating fragile single point of failure with limited redundancy. To address this issue, we propose distributed safety alignment (DSA), which redundantly encodes safety capabilities across multiple computational neurons, ensuring that the model maintains its safety baseline even when critical safety neurons are disrupted. Specifically, we localize the intervention to the inputs of the down-projection layers in language-side feed-forward networks and treat each feature coordinate as the activation of an individual neuron. DSA then combines neuron activations with loss gradients to compute a direction-aware first-order Taylor score that globally identifies the neurons that contribute most to the current refusal behavior of the model. Finally, targeted disruption via deterministic masking and stochastic dropout is coupled, forcing the model to abandon narrow safety neurons and redundantly encode safety behavior across multiple compensatory neurons. Extensive experiments show that DSA substantially improves robustness against white-box neuron-level safety attacks while preserving the model's general language and multimodal utility.

Span-guided rewriting aims to preserve meaning by localizing edits to annotated harmful spans, but the same constraint can leave harmful intent insufficiently mitigated. We present a controlled exploratory comparison of span-guided and unguided detoxification on a mixed-source English evaluation set comprising manually curated inputs and HateXplain test items. We conduct a dense blinded human evaluation under a fixed single-generator setting. Human preferences reveal a trade-off rather than a uniformly superior rewriting strategy. Span-guided outputs are favored when localized editing preserves the original stance and avoids unnecessary modification, whereas unguided outputs are favored when broader rewriting achieves more complete mitigation. This contrast varies substantially across the study-defined strata: the two strategies are competitive in the strong stratum, while unguided rewriting is clearly preferred in the mild stratum. Rationale annotations trace this difference to complementary failure risks: residual harm after localized editing and over-modification after broader rewriting. We treat automatic evaluation as a diagnostic rather than a substitute for human judgment. Toxicity-similarity scalarizations, a multi-generator analysis, and two general-purpose LLM judges reproduce parts of the aggregate tendency but do not yield an analogous stratified contrast. These setting-specific findings do not establish a severity-based routing rule. Instead, they motivate evaluation protocols that assess mitigation sufficiency and meaning preservation separately and report both residual harm and over-modification alongside aggregate scores.

Concept erasure techniques (CETs) edit text-to-image diffusion models to erase undesired targets such as NSFW content or copyrighted styles, while preserving model utility on benign concepts. Current CETs face a trade-off between erasure robustness and utility: stronger edits erase the target more reliably but degrade utility on non-target concepts, and vice versa. This stems from how existing methods define what to erase and what to preserve. Many CETs rely on static concept banks specified manually, generated by LLMs, or selected by CLIP image-text similarity. Such banks do not model how prompts steer the model during denoising, leaving it vulnerable to triggers that reintroduce the target while suppressing nearby benign concepts. We present Preservation-aware Adaptive Ranked Subspace Expansion (PARSE), a training-free framework for robust concept erasure in latent diffusion models. Given a target, PARSE queries the diffusion model with classifier-free guidance to dynamically discover target-inducing erase concepts and nearby retain concepts in the model vocabulary. It then edits the cross-attention value space with a preservation-aware projection that removes target directions while leaving retain directions intact. For triggers beyond this vocabulary-indexed space, PARSE iteratively searches for re-emergence triggers by textual inversion and adaptively expands the erased subspace only when a new trigger direction does not conflict with retain semantics. We also introduce the Balanced Erasure Utility Score (BEUS), which combines robustness (ASR under multiple attacks) and utility preservation (FID) via bounded monotone transforms and harmonic mean aggregation. Experiments on NSFW, artistic style, and object erasure, with a large-scale robustness-utility analysis over many CET baselines, show that PARSE erases multiple concepts robustly without sacrificing post-edit utility.

Span-level rationales are often assumed to improve controllability in text detoxification, but it remains unclear when such guidance helps and when it introduces trade-offs. We present Awareness-Enhanced Guidance for Iterative Safeguard (AEGIS) as an exploratory framework for studying span-guided multilingual detoxification across English, Mandarin Chinese, and Korean. AEGIS combines span-level detector outputs with frozen generator backbones, allowing harmful spans, intensity labels, and target attributes to be provided as structured guidance during rewriting. Rather than claiming state-of-the-art detoxification performance, we analyze how span guidance affects the balance between toxicity reduction and meaning preservation across generator families, model scales, and languages. Our results suggest that span-guided detoxification is conditionally useful: explicit rationales change the trade-off between toxicity reduction and meaning preservation, but their effects depend strongly on the generator backbone and the linguistic context. These findings highlight both the promise and the limitations of span-level control signals for multilingual detoxification.

Filesystem isolation in container ecosystems is often weakened by cross-boundary path misresolution, causing path traversal (PaTra) vulnerabilities. These vulnerabilities stem from insecure host-container interactions and have become increasingly pervasive as cloud systems mount shared resources, such as GPUs and agent workspaces, into containers to support AI workloads. Existing defenses remain inadequate. Kernel-level protections are intrusive, can destabilize system calls, and have therefore not been accepted into the Linux mainline. Detection methods rely on static rule matching or manual code auditing. Static rules can flag path-related functions but fail to capture the semantics needed to determine whether a host-container interaction exists, causing many false positives. Manual review requires domain expertise, making it costly, inefficient, and difficult to scale. To address this threat, we present Bulkhead, an automated framework that integrates large language models (LLMs) with formal methods for semantic vulnerability discovery and remediation. Bulkhead uses a multi-agent system to identify and repair PaTra vulnerabilities through multi-dimensional knowledge patterns generalized from known cases. It first applies high-risk functional patterns to locate entry points for cross-boundary interactions in containerized code, then uses call-chain patterns to recover the corresponding execution paths at suitable depth. The Detection pipeline analyzes these call chains against the application scenarios and threat model, identifying vulnerabilities such as missing security checks and TOCTOU flaws in cross-boundary interactions, and generating proof-of-concept (PoC) exploits for validation. These PoCs then guide patch generation. To ensure remediation correctness, the Patch pipeline performs assertion-driven verification using predefined model-checking templates.

Large Language Model (LLM) services introduce a fundamental privacy challenge. Sensitive information may be inferred not only from explicit identifiers, such as names or phone numbers, but also from contextual associations among otherwise innocuous spans. Existing sanitizers typically assign privacy or utility signals to individual spans without explicitly modeling pairwise relationships among them. In this paper, we propose PromptGraph, a graph-guided prompt-sanitization approach for privacy-preserving LLM inference. PromptGraph estimates privacy leakage at the span level and utility-relevant contextual dependencies between pairs of spans. It represents each prompt as an attributed graph, in which nodes carry span-level privacy scores and edges encode contextual dependencies needed to preserve utility. The sanitization objective selects a protected span set that maximizes privacy gain while penalizing the loss of contextual dependencies. This formulation explicitly balances privacy and utility when contextual evidence is hidden. Protected spans are sanitized locally, and returned placeholders are restored only after passing local consistency checks. We conduct extensive experiments showing that PromptGraph achieves a more favorable balance between privacy and utility than prompt-privacy baselines.

Open Radio Access Networks (O-RAN) increasingly delegate near-real-time control to deep reinforcement learning (DRL) xApps obtained from third-party vendors, creating a new supply-chain attack surface. A backdoor policy behaves optimally until an adversary injects a covert trigger into the observed key performance indicator (KPI) telemetry, at which point it issues harmful control actions that degrade quality of service (QoS). We present ORAN-DEFEND, a retraining-free wrapper that sanitizes a frozen, potentially compromised xApp by projecting each KPI window onto a safe subspace estimated from a small number of trusted clean rollouts via singular value decomposition (SVD). We establish, both analytically and empirically, a precise recovery condition: the defense succeeds if the trigger energy concentrates in the orthogonal complement of the safe subspace, and we quantify this boundary through the trigger's $\Eperp$ energy fraction. On the Colosseum COLORAN dataset, we evaluate four structurally distinct DRL backdoor attacks, like TrojDRL, SleeperNets, BadRL, and Q-Incept, spanning inner-loop and outer-loop poisoning regimes and demonstrate $100\%$ return recovery and $\geq99.5\%$ defense success rate across all four when the subspace assumption holds. A geometry ablation reveals an intrinsic and previously uncharacterized limit of any linear projection defense: when the trigger collocates with the legitimate signal, the $\Eperp$ energy fraction governs recovery monotonically, and the linear residual detector collapses to chance even while a nonlinear classifier retains perfect separability.

Jailbreak attacks remain a critical threat to the safe deployment of large language models (LLMs). While prior work has primarily studied attacks and defenses at the prompt level, we show that this prompt-centric paradigm overlooks a structural vulnerability in stateful, function-calling environments. In such applications, developer-defined schemas, structured arguments, and untrusted tool outputs are interleaved into a single shared model context. This architecture expands the attack surface by blurring the boundary between trusted control logic and untrusted data, allowing adversarial intent to be distributed across a multi-turn execution path. We exploit this architectural flaw through SMT, a black-box attack framework based on Simulated Moderation Traces. Departing from purely prompt-based interactions, SMT constructs a multi-turn trajectory that simulates a legitimate moderation-auditing workflow. Within this trajectory, a fabricated moderation frame leverages red-team testing as a pretext to elicit harmful generations. The subsequent validation feedback treats safety refusals as execution failures, prompting refinements that gradually weaken the model's safety constraints and ultimately trigger harmful outputs. Extensive empirical evaluations on prominent commercial LLMs from five different providers across two standardized safety benchmarks show that SMT consistently achieves the highest average attack success rate and HarmScore while requiring a near-minimal number of queries, substantially outperforming existing baselines. These findings demonstrate that prompt-level sanitization alone is fundamentally insufficient for defending tool-enabled LLM systems and highlight the urgent need for context-aware validation across schemas, arguments, tool outputs, and accumulated conversation state. The code is available at https://github.com/liujlong27/SMT.

Federated Learning (FL) enables collaborative model training without sharing raw data, making it a promising paradigm for privacy-sensitive applications. However, its decentralized nature makes it inherently vulnerable to backdoor attacks, where malicious clients embed hidden triggers into local training data to manipulate model predictions. Existing defenses mainly operate during before and during aggregation cannot fully eliminate backdoor behaviors that persist in the converged global model. Moreover, the effectiveness of post-training sanitization is often limited by the server's lack of knowledge of trigger patterns or poisoned clients after convergence, resulting in residual backdoor behaviors or accuracy degradation due to neuron entanglement. To address this limitation, we propose SCRUB-FL (Sanitizing and Cleansing Representations via Unlearning of Backdoors), a two-phase solution for post-training backdoor removal in FL. During training, clients identify suspicious samples using spectral analysis and activation clustering, then train lightweight Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP) models to capture trigger-related distributions. The generator parameters are aggregated server-side to construct a global representation of suspicious patterns without exposing raw data. After convergence, the server synthesizes trigger-approximating samples and applies machine unlearning to erase the trigger-target association by redistributing predictions toward a uniform distribution. Experimental evaluations on CIFAR-10 and GTSRB across three attack types and up to 40% malicious participation demonstrate that SCRUB-FL reduces the backdoor attack success rate to as low as 3.88% while maintaining over 91% normal task accuracy, outperforming state-of-the-art defenses without requiring prior trigger knowledge or a large clean proxy dataset at the server.

Large language model applications build prompts from templates, and Handlebars is a widely used templating engine and the default prompt-template format in Microsoft Semantic Kernel. Its double-brace {x} expression HTML-escapes the interpolated value and is documented as the safe default; its triple-brace {x} expression inserts the value raw. We show that this choice silently governs an application's exposure to structural role injection, where attacker-controlled data carries chat role delimiters that forge a higher-privilege turn. A model-free analysis establishes the mechanism: Handlebars escaping rewrites angle brackets but not square brackets, colons, or Markdown hashes, so it neutralises ChatML, Llama-3, and XML role delimiters (survival rate 0.00) while leaving Llama-2 [INST], legacy Human:/Assistant:, and Markdown ### delimiters intact (survival rate 1.00 for the last two). We then run 5760 trials across seven delimiter families, two attack objectives, and four models (GPT-3.5 Turbo, GPT-4o mini, GPT-4.1 mini, Claude Haiku 4.5) at a combined API cost of 1.63 USD. GPT-3.5 Turbo follows the task-hijack instruction in 97% of raw and 91% of escaped trials, with the escaping protection concentrated in the angle-bracket families and absent for the colon- and Markdown-based families; the harder secret-exfiltration objective, which does not saturate, exposes the same family interaction more cleanly. Claude Haiku 4.5 resists both objectives almost entirely. The escaped default protects only the delimiter schemes whose characters HTML escaping happens to cover, gives no protection for the rest, and cannot substitute for a structural separation of instruction and data.

Prompt injection defenses evaluated on synthetic benchmarks do not generalize to real enterprise documents, which are longer, denser, and interleave legitimate authority language with factual content. We demonstrate this gap with a real-document benchmark of 122 tasks across five professional domains (financial, legal, medical, scientific, DevOps) using actual SEC filings, Federal Register rules, PubMed abstracts, arXiv papers, and GitHub postmortems. Paraphrasing, the strongest defense on synthetic benchmarks, shows no statistically significant attack success rate reduction on real documents (p=0.500) while degrading utility from 91.8% to 82.8%. We introduce PARSE (Provenance-Aware Retrieval Sanitization), a domain-aware, fact-preserving sanitization pipeline that classifies each sentence by injection likelihood, extracts structured facts before rewriting, and verifies fact preservation via a consistency-checking loop. A directiveness gate routes 59% of real enterprise documents to a lightweight path, concentrating computational cost on high-risk documents. PARSE achieves 15.6% attack success rate -- a 38% reduction versus the 25.4% baseline -- at 86.9% utility, the only condition that is both statistically significant (p=0.014, adequately powered) and maintains near-baseline utility. Practitioners should evaluate defenses on domain-matched real documents, not synthetic proxies.

When distributed agents exchange text across organizational boundaries, privacy leakage arises not only from explicit identifiers but also from distributional signatures such as formatting conventions, vocabulary choices, and syntactic patterns. We propose DiSan(Disentangled Sanitization), a privacy-preserving sanitization framework and a built-in component of Intern-Shannon for multi-agent collaboration. DiSan uses a two-stream encoder to factorize text into a source-invariant role subspace that preserves task semantics and a source-identifying style subspace that remains local. Federated proto-type alignment and adversarial regularization enable joint training without centralizing raw text. Experiments show that identifier-level masking is insufficient: masking 19.2% of tokens reduces TF-IDF stylometric attribution by only 18.6%. By contrast, DiSan reduces answer-level PII exposure by 20 times while maintaining 83% answer faithfulness on a distributed multi-agent RAG benchmark, and lowers Enron stylometric attribution by 73.2% under TF-IDF and 70.6% under a neural probe.

Modern LLM-powered autonomous agents increasingly rely on rich user interface (UI) state observations to achieve reliable action grounding in complex digital environments. However, many deployments transmit the full UI state to remote inference servers even when most elements are irrelevant to the current task, which can leak sensitive but unnecessary context such as authentication codes, private notifications, and background application states. We propose MINIM, a trusted local broker that performs privacy-aware minimization on the client side before any observation leaves the device. Grounded in Contextual Integrity (CI), MINIM learns a dual-score representation for each UI element by predicting an inherent sensitivity score (s) and a task-conditioned necessity score (n). These scores drive a ternary disclosure policy that keeps essential elements, abstracts sensitive attributes when needed, and removes task-irrelevant content. We optimize a CI-aware objective that penalizes necessity errors more strongly on high-risk content, enabling aggressive pruning while preserving task-critical information. Experiments on real-world UI observations derived from WebArena show that MINIM substantially reduces task-irrelevant sensitive leakage while preserving task-critical semantic context and the interactive affordances required for reliable agent actions.

Multimodal large language models (MLLMs) have raised new privacy challenges. On the data side, user-provided inputs often include unpredictable sensitive information; while on the downstream task side, model reasoning depends on rich visual context that may itself be privacy-sensitive. Existing privacy protection methods, however, rely on predefined sensitive categories and fixed obfuscation strategies, struggling to tackle such challenges in MLLMs. To address this dilemma, we propose Anchored Privacy Drifting (APD), a training-free method that drifts privacy-sensitive elements toward semantically equivalent alternatives while anchoring contextual cues to the source image. To systematically evaluate this dual objective of privacy protection and contextual preservation, we introduce AdaptShield, a comprehensive benchmark covering 22 privacy categories, which combines conventional privacy metrics with MLLM-based assessments of contextual utility. Extensive experiments show that our method achieves balanced improvements in both privacy sanitization and content retention, with average gains of 10.4% on textual categories and 8.5% under MLLM-based evaluation across four MLLM series, i.e., Qwen2.5, Qwen3, InternVL3, and InternVL3.5.

Audio and omni-modal large language models exhibit impressive cross-modal reasoning capabilities. However, applying standard reinforcement learning post-training algorithms to these models exposes a critical structural vulnerability: methods like GRPO apply uniform policy gradients across all tokens, ignoring their unequal dependence on the non-text source modality. This exacerbates late-stage modality collapse during extended chain-of-thought generation, where models progressively abandon the primary source signal in favor of compressed textual priors, leading to confident but ungrounded hallucinations. To address this, we introduce Modality-Aware Policy Optimization (MAPO), a novel dual-branch reinforcement learning framework. First, MAPO dynamically concentrates the policy gradient on modality-critical tokens using a modality relevance mask, which is derived from the cross-modal differential entropy between an audio-ablated reference and the multimodal policy. Second, it integrates an auxiliary attention loss branch that applies a targeted, temporally scaled penalty to the model's internal attention distributions. This ensures the model actively sustains cross-modal grounding deep into the reasoning trace. Evaluations on complex audio reasoning benchmarks demonstrate that MAPO substantially improves long-horizon reasoning fidelity and multimodal instruction following, achieving highly competitive performance and setting new state-of-the-art results on several key benchmarks among open-weight models. By relying strictly on native statistical signals rather than domain-specific inductive biases, MAPO offers a promising foundation for mitigating epistemic collapse across diverse multimodal systems.

We study exact constant collapse in variational autoencoders: the deterministic encoder mean becomes independent of the input. The prior remains the standard Gaussian. Before VAE training, we select a fixed teacher posterior from a GMM-based view of the data and attach a fixed latent-only simplex witness to the encoder mean. This construction yields two linked objects. The first is a certificate: if the witness prediction improves on the best constant predictor of the teacher, the encoder mean cannot be input-independent constant. The second is a local escape direction: on the collapsed manifold, the teacher residual gives a sample-dependent descent direction for the alignment loss. For any full-support teacher posterior, the same geometry also gives a closed-form latent code with zero teacher-witness alignment error. Its scaled versions trace a margin-energy path from the constant predictor to the exact teacher code, which quantifies non-collapse inside the protected witness subspace. We instantiate the method on MNIST, CIFAR-10, and CIFAR-100. With searched unsupervised PCA-GMM teachers, vanilla VAEs fail the teacher-witness certificate in all five seeds on CIFAR-10 and CIFAR-100, while RST variants pass in all five seeds. Under collapse-stress settings with $β_{\mathrm{KL}}\in\{2,4,8\}$, vanilla VAE again fails in all seeds, whereas RST-alpha-prefit remains certificate-positive. Escape trajectories on both natural-image datasets increase the witness margin from a low-margin initialization and exhibit nonzero teacher-induced gradient norms. The analysis is confined to exact constant collapse of the encoder mean; generation quality, decoder use, and other collapse modes remain separate questions.

An agent's probability report is paid for twice: by a strictly proper scoring rule, and by an approval rule for the decision it triggers. In this classical decision-coupled setting, non-affine approval is known to defeat truthful reporting. We show the conflict is endogenous: when feasible, the welfare-maximizing approval rule is never affine. The distortion, however, is predictable and can be designed around. There is a reserve report at which pretending to be the marginal type costs exactly the approval prize. Approving at or above the reserve screens types perfectly under every strictly proper score, and the reserve does not depend on the type distribution. A Lipschitz rule with a single kink attains first-best exactly; under strict feasibility no continuously differentiable rule does. The binding constraint is steepness, not smoothness. First-best is attainable within a slope budget if and only if the budget is at least the critical slope: the steepest chord of the pretending cost up to the reserve. Below it the welfare loss is cubic in the shortfall. Where the pretending cost is convex up to the reserve, as for Brier, log and power scores, the critical slope is closed-form. The instances are AI-agent oversight and marketplace operation.

Retrieval-augmented generation and tool-integrated LLM agents increasingly depend on external textual sources. This reliance broadens the available attack surface, allowing adversaries to insert malicious instructions that trigger unintended model behaviors. Current defensive measures often utilize LLM-based detectors to filter such content, but these approaches remain vulnerable to optimization-based attacks. Additionally, training-based methods frequently fail to generalize to novel data distributions. To resolve these issues, we introduce SONAR, a prompt sanitization framework that identifies and removes injected content using metrics from natural language inference. Specifically, SONAR constructs a sentence-level relational graph across the user query and external data. By using entailment and contradiction scores as edge weights, the system identifies sentences that deviate from the core task. It then employs connectivity-driven pruning to eliminate flagged injection seeds and their related neighbors while maintaining benign context. Rigorous evaluations across several models and datasets show that SONAR reduces the attack success rate to nearly zero, significantly outperforming nine established baseline defenses.

Large language models (LLMs) increasingly act as autonomous agents, using tools to execute code, read and write files, and access networks, creating novel security risks. To mitigate these risks, agents are commonly deployed and evaluated in isolated "sandbox" environments, often implemented using Docker/OCI containers. We introduce SANDBOXESCAPEBENCH, an open benchmark that safely measures an LLM's capacity to break out of these sandboxes. The benchmark is implemented as an Inspect AI Capture the Flag (CTF) evaluation utilising a nested sandbox architecture with the outer layer containing the flag and no known vulnerabilities. Following a threat model of a motivated adversarial agent with shell access inside a container, SANDBOXESCAPEBENCH covers a spectrum of sandboxescape mechanisms spanning misconfiguration, privilege allocation mistakes, kernel flaws, and runtime/orchestration weaknesses. We find that, when vulnerabilities are added, LLMs are able to identify and exploit them, showing that use of evaluation like SANDBOXESCAPEBENCH is needed to ensure sandboxing continues to provide the encapsulation needed for highly-capable models.

Unlike regular tokens derived from existing text corpora, special tokens are artificially created to annotate structured conversations during the fine-tuning process of Large Language Models (LLMs). Serving as metadata of training data, these tokens play a crucial role in instructing LLMs to generate coherent and context-aware responses. We demonstrate that special tokens can be exploited to construct four attack primitives, with which malicious users can reliably bypass the internal safety alignment of online LLM services and circumvent state-of-the-art (SOTA) external content moderation systems simultaneously. Moreover, we found that addressing this threat is challenging, as aggressive defense mechanisms-such as input sanitization by removing special tokens entirely, as suggested in academia-are less effective than anticipated. This is because such defense can be evaded when the special tokens are replaced by regular ones with high semantic similarity within the tokenizer's embedding space. We systemically evaluated our method, named MetaBreak, on both lab environment and commercial LLM platforms. Our approach achieves jailbreak rates comparable to SOTA prompt-engineering-based solutions when no content moderation is deployed. However, when there is content moderation, MetaBreak outperforms SOTA solutions PAP and GPTFuzzer by 11.6% and 34.8%, respectively. Finally, since MetaBreak employs a fundamentally different strategy from prompt engineering, the two approaches can work synergistically. Notably, empowering MetaBreak on PAP and GPTFuzzer boosts jailbreak rates by 24.3% and 20.2%, respectively.

The integration of Large Language Models into daily tasks relies on context-rich instructions, inevitably exposing sensitive user information. Current privacy-preserving methods typically employ context-agnostic static rules, causing severe utility degradation. However, the specific mechanisms governing how sanitization impacts downstream performance remain largely underexplored. To address this, we conduct a systematic analysis to deconstruct the privacy-utility trade-off, uncovering three underlying mechanisms: (1) Context-Dependent Utility, which first establishes when to sanitize by revealing that data value shifts from critical constraints to dispensable noise based on user intent; (2) Strategic Adaptation, which subsequently determines how to sanitize by dictating that the choice between removal and replacement depends on the task's reliance on factual integrity versus structural coherence; and (3) Combinatorial Interplay, which finally extends the protection scope by demonstrating that attributes form a semantic web of synergistic dependencies or antagonistic redundancies. Guided by these insights, we introduce an intent-driven local protection framework. By distilling a lightweight model Veilmind-4B to drive a dynamic extraction-sanitization-restoration pipeline, our approach reaches a low-leakage privacy point while preserving substantially higher response utility than existing privacy-oriented baselines, advancing the privacy-utility trade-off toward the Pareto frontier.