Applications such as personalized assistance and proprietary document analysis require large language models (LLMs) to generate outputs from private data. Yet powerful LLMs typically cannot be deployed on the resource-constrained devices where private data resides, and uploading private data to cloud-hosted LLMs exposes sensitive information. Recent work addresses this tension with a cloud-edge collaborative decoding paradigm, where private data are kept on the edge with a small language model (SLM) producing next-token distributions, which are fused with predictions from a cloud LLM operating solely on public data. In this paper, we systematically analyze the privacy risks of such a paradigm with a novel evaluation framework using constructed QA datasets, which show that such collaboration can expose substantial private-context information. To address such privacy leakage, we propose CoVeil, a defense mechanism which dynamically optimizes transmitted signals to suppress leakage during decoding time while preserving the collaborative quality. Extensive evaluations demonstrate that CoVeil consistently improves the privacy-utility trade-off over existing baselines by reducing data leakage by up to 87.2%, with minimal accuracy loss.
Multi-agent LLM systems often coordinate by compressing an upstream interaction into a handoff artifact that downstream agents treat as shared state. We show that this handoff step is a structural source of privacy leakage: summaries preferentially preserve operational facts while weakening the boundary metadata that governs how those facts may be used---a failure mode we call \emph{summary collapse}. On a controlled multi-agent coordination testbed we measure marker survival with a human-validated judge ($κ= 0.74$), where $σ_b = 1$ means every boundary marker survives verbatim and $σ_b = 0$ means all are lost. Boundary-marker and operational-fact survival are nearly uncorrelated at the handoff level on both GPT-5-mini and DeepSeek-R1-32B (Pearson $r$ near zero): uncompressed free-text handoffs preserve boundaries at $σ_b \approx 0.80$, whereas a $25$-word budget drops $σ_b$ to ${\approx}0.57$ while operational-fact survival stays near ceiling. Controlled downstream tests reveal that protection depends on \emph{boundary explicitness}: vague languages leak in $73\%$ of GPT and $50\%$ of DeepSeek cases, while explicit constraints reduce leakage to under $15\%$ across all three tested models. A no-handoff single-agent control further shows the failure is not reducible to multi-agent topology as direct full-marker access still leaks more often than the operationalized handoff. Prompt-only mitigation and exact-string redaction only partially address the problem, while a gold-derived audience allowlist nearly eliminates leakage across models, showing that correctly identifying audience boundaries is the key factor.
Reinforcement learning with verifiable rewards (RLVR) is deployed to make models better at reasoning tasks, but its side effect on what models will divulge is under studied. Here we show that RLVR on facts increases extraction of personally identifiable information (PII) the instruct model had already memorized. We first confirm that instruct models have already memorized PII but leave them latent, rarely surfacing one when asked. We then apply RL on benign factual data that contains no PII of any kind, and re-probe: a targeted probe over name->email pairs, and an untargeted free-recall prompt that simply asks the model to list the addresses it knows. PII extraction rises sharply under both: on DeepSeek-V3.1, verbatim recall@k increases from 0.155 to 0.370, a 2.4x gain. The effect scales with model size: across three models spanning 8B to 671B parameters, absolute leakage is largest in the biggest model. Meanwhile model's reasoning abilities and refusal rates are retained, indicating that RL selectively changes which memorized information is accessible rather than broadly altering the model. In summary, memorized private data can be made markedly more extractable by training that never touches it. This gives an adversary a route to memorized data that requires no privacy-relevant training signal and no access to the data itself -- only the ability to fine-tune on something innocuous.
For AI agents to be useful beyond simple chat, they must hold sensitive user context such as calendars, credentials, health records, and financial data. We study whether the mere presence of such secrets in a model's context window introduces hidden correlations into the model's benign outputs, allowing reconstruction even when the model correctly refuses direct extraction. We further study whether an adversary can actively engineer prompts that amplify this effect, using the model as a covert carrier to transmit secrets through seemingly innocuous text. In both cases, this limited leakage is exploited using a novel adaptive attack that assumes black-box access to the underlying model. In controlled experiments across eight proprietary models, we find that 2-digit in-context secrets are reconstructed with near-perfect accuracy and 4-digit secrets at 82\% exact match, all from outputs the model produces in response to ordinary, non-adversarial requests. We observe that more capable models leak more: stronger instruction-following amplifies sensitivity to in-context secrets, suggesting leakage is a byproduct of capability as opposed to a patchable bug. We show this leakage enables two practical attacks: (1) a trained classifier that infers semantic predicates about user memories (e.g., health conditions, financial events) from routine natural-language outputs, and (2) an RL-trained adversary that extracts full Social Security Numbers from a production-style agent.
Beining Xu, Hairui Wang, Jiaxin Wang +2cs.CV cs.CR cs.MM
While the privacy risks of multimodal large language models (MLLMs) have drawn significant attention, the unique vulnerabilities of domain-specific MLLMs remain largely underexplored. Focusing on document understanding MLLMs for identity document processing, this paper investigates the privacy issues inherent in Key Information Extraction (KIE) tasks. We reveal that when input images lack sufficient visual evidence, these models often rely on memorized field relations from training data to infer missing content, thereby leaking multiple correlated fields containing sensitive personal information. To mitigate this risk, we make three key contributions.First, we propose the Dynamic Relational Unlearning Framework (DRUF) which comprises a Relational Decoupling Unlearning (RDU) module and a dynamic set update mechanism. It suppresses the leakage of high-risk field pairs while preserving KIE performance.Second, we introduce DocPrivacyBench, a novel benchmark to systematically evaluate a model's susceptibility to privacy leakage under conditions of absent or minimal visual evidence.Third, we evaluate three MLLMs and six unlearning methods using this benchmark, assessing both post-unlearning leakage suppression and utility preservation.Our results demonstrate that existing MLLMs consistently exhibit privacy leakage when visual evidence is scarce, particularly on noisier datasets. In contrast, DRUF outperforms the strongest baseline by improving leakage suppression by 4.8 percentage points, effectively mitigating privacy risks while maintaining robust document information extraction performance.
Reinforcement-learning (RL) post-training equips multimodal large reasoning models (MLRMs) with exploratory chains of thought (CoT), substantially improving visual reasoning. However, we find that this capability introduces a distinct privacy vulnerability: even when a sensitive fact is successfully unlearned from the final answer, the model may still reproduce it in its reasoning trace. This leakage is substantially more pronounced in natively RL-trained MLRMs than in their non -reasoning base models, revealing a privacy risk that existing unlearning methods are not designed to address. We show that RL-induced exploration leaves sensitive content with a distinctive token-level entropy signature that is largely absent from base models. Based on this observation, we propose LEMUR, a fully training-free, inference-time unlearning framework for natively RL-trained multimodal models. LEMUR uses entropy dynamics as a control signal to identify when sensitive reasoning begins and when sanitization should stop. During this interval, it redirects the reasoning trajectory through entropy-modulated visual-anchor latent injection, replacing committed tokens with sanitized, probability-weighted embeddings re-grounded in the input image. Across diverse MLRMs, LEMUR consistently outperforms existing unlearning met hods in suppressing both reasoning-trace and answer leakage, while better preserving non-sensitive utility and output fluency. These results demonstrate that RL-induced entropy dynamics provide a distinctive signal for privacy leakage and that exploiting this signal enables effective training-free unlearning for reasoning-capable multimodal models.
Yongli Xiang, Zhifang Zhang, Bojun Yang +4cs.CR cs.CL cs.CY
Persona skills distill personal interaction histories into portable and executable artifacts for downstream agents. While enabling flexible personalization, this process concentrates fragmented personal signals, amplifies their impact through reuse, and challenges defenses designed for individual records or retrieval-based memory. To systematically investigate the safety of the persona-skill pipeline, we introduce AntiSkillBench, an end-to-end benchmark for evaluating risks and defenses across the persona-skill pipeline. It comprises: (i) a dataset of 7,500 persona-grounded dialogue traces, constructed from 50 behaviorally rich profiles spanning diverse task scenarios; (ii) an evaluation suite that measures skill-level privacy leakage and agent-level attribute disclosure and behavioral impersonation across three skill-distillation strategies; and (iii) a defense evaluation covering four configurations across online and post-hoc interventions, including active risk suppression and passive provenance protection. Experiments across three frontier agents show that persona-skill risks persist across agent backbones and distillation protocols, extending from explicit attributes to communication styles and personality traits. Existing defenses exhibit limited and distillation-dependent effectiveness, failing to generalize across risk and distillation strategies. These results highlight AntiSkillBench as a challenging benchmark for developing privacy-preserving and authenticity-aware persona skills.
Jong Wook Kim, Byoungjae Min, Kennedy Edemacu +3cs.CR cs.CL cs.LG
Long-term memory enables persistent personalization in LLM agents, but repeated memory-conditioned responses can cumulatively reveal protected attributes even when they are never stated explicitly. We formalize this threat as adaptive transcript privacy and introduce DP-MemView, a differentially private interface that privately selects public response-conditioning views and exposes those views---rather than raw memory---to the response LLM. Each private selection is charged to every protected attribute whose memory group intersects the read set. Per-attribute ledgers block any selection that would exceed its cap and return a fixed generic view instead. Under an explicit interface contract, we prove pure B_a-DP for the entire adaptive transcript. We also extend the result to stores that differ across multiple protected groups and bound how much observing the transcript can change an adversary's prior odds. We evaluate the online and preallocated modes with three response LLMs on a controlled adjacent-store benchmark and a public-corpus transfer track. Both modes keep transcript distinguishability near chance while preserving target-required personalization and overall response quality. Further diagnostics show that removing key safeguards causes mismatched output support, missing ledger charges, revealing side channels, or growing long-horizon leakage.
A LoRA adapter is a few megabytes that almost everyone treats as a skill rather than a record of the data behind it. We put that assumption on a scale. Extending compression-based memorization analysis to the frozen-base setting, we measure directly, in bits, how much a low-rank adapter writes into a model it never changes. The answer is both smaller than full fine-tuning and less lawful than parameter counting would predict. Adapters store a couple of bits per trainable parameter, well short of a full model's budget, but that figure turns less on how many parameters an adapter carries than on where they sit. Move the same parameter budget from attention into the MLP and it holds nearly twice as much; strip the frozen base of its structure and the capacity all but disappears. Applied to realistic fine-tunes of Qwen2.5, the same instrument shows privacy leakage rising with the bits an adapter writes rather than the parameters it nominally has, and it draws a clean line between supervised and reinforcement learning: the secrets that supervised fine-tuning copies down verbatim, an adapter trained on verifiable rewards never records. Measuring what fine-tuning writes, rather than attacking it after the fact, turns a piece of folklore into a quantity one can design against.
Ryan Thornton, Mir Mehedi Ahsan Pritom, Maanak Guptacs.AI cs.CR
Large Language Models (LLMs) are capable of generalizing human language for the completion of never-before-seen tasks, leading to widespread deployment. While this automation provides clear utility, completing these tasks often requires the insertion of Personally Identifiable Information (PII), strings of information that uniquely identify some individual, raising privacy concerns. However, ethics has prevented the curation of a public, authentic dataset of PII. Without an appropriate dataset, it is difficult to quantify privacy risks. Thus, we introduce the PANOPTICON pipeline and dataset. The dataset, generated by Meta's Llama-3.1-8B-Instruct model, contains 67, 718 prompts, intended for the models context window, containing PII spans derived from 9,674 publicly available synthetic user profiles. We measure lexical diversity and S-BERT diversity of the created dataset to evaluate realism. Finally, we present a case study showcasing the utility of PANOPTICON data for understanding Prompt Inversion Attacks (PIAs). PANOPTICON thus emerges as the first benchmark dataset for studying PIAs over private corpora, providing a foundation for future LLM privacy research.
The flourishing code hosting platforms and coding agents enable even beginners with private data to build tailored Machine Learning (ML) models using available code quickly. The training data for ML models, often regarded as private property (e.g., clinical records, transaction information), is at significant risk of information leakage. Property Inference Attacks (PIAs), as a significant type of privacy attack, aim to expose global property information of the training set. In this paper, we present Code-Poisoning Property Inference Attack (CPPIA), the first code-level PIA, which overcomes four limitations of existing works: insufficient attack performance, severe degradation of model accuracy, high computational overhead, and failure under defenses. We consider malicious code providers from code hosting platforms (GitHub) and coding agents (Codex). Upon downloading the poisoned code, data holders train models with their private data without professional auditing, subsequently releasing label-only APIs to the public. The adversary embeds the properties into secret samples during training and queries the trained model on these samples later to leak privacy. CPPIA offers 100\% attack accuracy without degrading model accuracy. It is also computationally lightweight and requires no shadow models. We evaluate the attack performance across four datasets, eight model architectures, eighteen properties, and under three defense mechanisms, demonstrating the universality and effectiveness of CPPIA.
Santhosh Parampottupadam, Andres Martinez, Dimitrios Bounias +3cs.LG cs.CL cs.CR
Federated learning (FL) enables multi-institutional training on clinical text without sharing raw data, but gradient inversion can reconstruct sensitive information from shared model updates. The extent of this leakage for radiology reports, and the role of tokenizer design, remains unclear. We quantify gradient-based text reconstruction in FL and compare privacy risk across three tokenizers with the model architecture held fixed. Six FL clients trained a GPT-2-style transformer (sequence length 32) on public radiology corpora (368,751 diagnostic reports, 98,206 discharge summaries, 1,500 MIMIC-CXR free-text reports) using the GPT-2, RadBERT, and LLaMA-2 tokenizers at batch sizes of 64, 128, and 256. Assuming an active malicious server that modifies the shared architecture before distribution, we applied analytic gradient inversion and measured reconstruction fidelity over five runs. Exact sentence reconstruction ranged from 31% to 44% across tokenizers (30.6-43.5% across the 27 tokenizer x dataset x batch-size cells). At batch size 64 on the Discharge dataset, accuracy was 42.1% (GPT-2), 42.3% (RadBERT), and 39.4% (LLaMA-2), decreasing to 37.3%, 37.2%, and 34.3% at batch size 256. S-BLEU declined as batch size grew (GPT-2: 0.44 to 0.33; RadBERT: 0.48 to 0.35). RadBERT yielded the highest reconstruction fidelity and recovered the most clinical terms (18.1% of a 1,440-term reference vocabulary, vs 12.5% for GPT-2 and 9.4% for LLaMA-2), yet no tokenizer prevented leakage. Substantial portions of report text are therefore recoverable from FL gradients even at larger batch sizes and with domain-specific tokenizers. Tokenizer design influences leakage severity and is a privacy-relevant decision, not only a utility one; safeguards such as secure aggregation and differential privacy are likely necessary to meet HIPAA and GDPR requirements for FL in radiology NLP.
To overcome data scarcity and privacy constraints in data collection, it has become standard practice across academia and industry to augment real training data with text-to-image (T2I)-generated synthetic data, a paradigm we term Real-Synthetic Mix-Training (RSMT). While substituting synthetic data for sensitive real samples is widely regarded as a means to mitigate privacy exposure of the substituted data, the risk to the remaining real samples that actively participate in training has remained largely unexamined. This work reveals, for the first time, that RSMT can substantially amplify privacy leakage of these real training samples. We establish a theoretical framework, RSMT Memorization Amplification, proving that incorporating synthetic data displaces real samples toward peripheral regions of the mixed feature space, in turn forcing the model to memorize them more aggressively. Guided by this foundation, we propose RSMixLeak to systematically assess this risk through membership inference attacks (MIAs). RSMixLeak comprises two variants depending on the adversary's capability. The non-adversarial variant audits a benign RSMT pipeline with an honest T2I provider, establishing a lower bound on the leakage induced by the intrinsic gap between real and T2I-generated data. The adversarial variant considers an adversary who controls the T2I model or contributes crafted data to the T2I provider, and deliberately enlarges this distributional gap on a target class via either high-level semantic attribute binding or imperceptible pixel-level coating, further amplifying leakage on real training data while improving downstream model utility. Motivated by these findings, we further propose a lightweight leakage propensity indicator computable from real data alone that reliably identifies high-risk datasets unsuitable for entering RSMT, as a self-assessable mitigation.
Large language models increasingly mediate interactions between sensitive data, untrusted inputs, and privileged actions in agentic systems, creating security and privacy risks. These range from prompt injections that manipulate downstream tool use to leakage of confidential information through model outputs. Recent Information Flow Control (IFC)-based defenses show promise but lack a principled semantic foundation for reasoning about information flow through the model itself. Since any input token may influence any output token in an autoregressive LLM, existing approaches suffer from severe taint explosion. We present Geometric Information Flow (GIF), a semantic framework for tracking information flow from input tokens to outputs. GIF uses the LLM Jacobian and local output geometry to upper-bound the Shannon mutual information between perturbed input spans and model outputs, yielding a scalable measure computable on large models via automatic differentiation and low-rank approximation. Unlike attention-based or correlational attribution heuristics, GIF satisfies local geometric soundness, and we provide a fully mechanized Lean 4 proof that it upper-bounds the true information flow induced by a given prompt under local regularity assumptions. We evaluate GIF on integrity and confidentiality tasks across multiple prompt-injection and privacy-leakage benchmarks. GIF achieves near-perfect recall even without a downstream declassifier, outperforming attention-based baselines. Combined with lightweight LLM-based declassifiers, it matches or exceeds the F1 of direct LLM-as-judge baselines such as GPT-5.5 xhigh reasoning while using up to 81x lower token cost. GIF flows detected with small surrogate models transfer to larger state-of-the-art models and other model families, even when the surrogate is up to 200x smaller, suggesting black-box deployment without gradient access.
Elena Sofia Ruzzetti, Cornelius Emde, Sangdoo Yun +2cs.CL cs.AI
LLM agents are increasingly deployed in multi-party environments, handling sensitive personal data on behalf of individual users, for instance in group chats. When such an agent discloses private information, it reaches every group member at once. This risk is structurally harder to control than in one-to-one settings, as every piece of private information must be appropriate for every recipient in the group. Yet all existing contextual privacy benchmarks consider only single-interlocutor settings, leaving multi-party privacy risks unmeasured. We introduce MuPPET (Multi-Party Privacy Exposure Testing), a benchmark for contextual privacy in multi-party conversations. Our experiments show that models leak substantially more in multi-party settings than one-to-one evaluations suggest. Frontier models are vulnerable, and smaller open-weights models, often preferred for local deployment with sensitive data, even more so. Existing contextual privacy defences offer only partial protection, degrade utility, and do not resolve the underlying party-tracking problem.
Moon Ye-Bin, Nam Hyeon-Woo, Baek Seong-Eun +2cs.CR cs.AI
Agents are increasingly deployed in document-intensive workflows where sensitive private information is not an edge case but a routine input, e.g., an agent booking a flight needs passport numbers. In such settings, the agent must use private information to complete tasks accurately while never exposing it in its responses, because it cannot verify who is actually at the keyboard. These two obligations are in fundamental tension. A model capable enough to use private information for task completion can, by the same capability, be induced to reveal it. To evaluate the trade-off of task accuracy and privacy leakage, we introduce Task-completion and Resistance to Active Privacy-extraction (TRAP). Each scenario includes a document containing private information, a task query that requires the agent to invoke the correct tool using private fields, and an attack query that attempts to elicit the same information in natural language. Evaluating 22 models spanning frontier proprietary and open-source models at multiple scales, we find that all model families exhibit non-trivial leakage, and that instruction-following ability correlates with leakage rate. Existing prompt-based defenses reduce leakage but at significant cost to task accuracy. Prompt optimization fails to escape this trade-off. We demonstrate that this failure is not incidental. For any softmax-based model, no soft-constraint defense, e.g., prompt-based defenses, can jointly achieve high task success with zero leakage probability. Motivated by this impossibility result, we propose structural private field isolation, which replaces private fields with hash keys before they reach the model. This approach largely prevents leakage while keeping task accuracy.
We ask whether stereotype-loaded queries about culturally marked people leak more personal information from a retrieval-augmented generation (RAG) system than otherwise-equivalent neutral queries. We pre-register a four-culture audit (en-Anglo, es-LATAM, Arabic, Hindi) on a synthetic English PII corpus, comparing five query arms we call the Stereotype-Trigger Leakage Delta (STLD). Two caveats up front. Our locked confirmatory estimator was never run, so every test in the paper is exploratory or sensitivity, with all plan deviations listed in the appendix. And the name-leakage metric is contaminated by a prompt-echo artifact: the model often just re-emits the name we asked about, which inflates apparent leakage without any retrieval at all. On the cleaner channels (email, phone, ssn-like, address), we find no stereotype-driven amplification on any of the four cultures after multiple-comparison correction. Because our sample is only powered for mid-sized effects, and because the culturally marked probes mix stereotype content with cultural markers and heritage practices, we present this as no detection, not evidence of no effect, of culturally marked predicate leakage that is confounded with the underlying resource.
Privacy risks in text-only Large Language Models (LLMs) are well studied, particularly their tendency to memorize and leak sensitive information. However, Multi-modal Large Language Models (MLLMs), which process both text and images, introduce unique privacy challenges that remain underexplored. Compared to text-only models, MLLMs can extract and expose sensitive information embedded in images, posing new privacy risks. We reveal that some MLLMs are susceptible to privacy breaches, leaking sensitive data embedded in images or stored in memory. Specifically, in this paper, we (1) introduce MM-Privacy, a comprehensive dataset designed to assess privacy risks across various multi-modal tasks and scenarios, where we define Disclosure Risks and Retention Risks. (2) systematically evaluate different MLLMs using MM-Privacy and demonstrate how models leak sensitive data across various tasks, and (3) provide additional insights into the role of task inconsistency in privacy risks, emphasizing the urgent need for mitigation strategies. Our findings highlight privacy concerns in MLLMs, underscoring the necessity of safeguards to prevent data exposure. Our dataset and code can be found here.
Zifan Peng, Yini Huang, Aiwen Lu +7cs.CR cs.AI cs.CY
Public social media posts can reveal private information through weak cues scattered across text, images, or metadata. Such leakage is often cumulative and cross-post: cues that appear harmless in isolation may jointly expose a user's home, workplace, or routine. However, current research lacks a unified benchmark for user-level multimodal privacy leakage and an evaluation metric that captures exposure severity beyond binary accuracy. To address these gaps, we propose SopriBench, a synthetic benchmark guided by leakage patterns abstracted from a private reference corpus of Rednote and Instagram accounts, covering 50 user profiles and 1,569 images with attributes, contextual sensitivity, granularity, leakage type, inference difficulty, and supporting evidence. We further introduce the Privacy Exposure Score (PES), which weights value granularity by contextual sensitivity. Inspired by abductive reasoning, we introduce Argus, a training-free agentic framework for cumulative leakage inference. Argus forms hypotheses from accumulated evidence, verifies supporting evidence, and aggregates cross-post cues into privacy profiles, achieving 0.55 PES, a 25% improvement over the strongest baseline, with the largest gain on cross-post leakage.