Privacy Leakage in Language Models

LLMs memorize sensitive training data, including personally identifiable information (PII), creating a pressing need for reliable post hoc removal methods. Unlearning has emerged as a promising solution, with state-of-the-art(SOTA) methods often following a localize-first, unlearn-second paradigm that targets specific model parameters. However, existing benchmarks evaluate unlearning solely at the output level, leaving open the question of whether unlearning truly erases knowledge from a model's parameters or merely obfuscates it, a concern reinforced by the success of resurfacing attacks. To bridge this gap, we introduce LACUNA: the first unlearning testbed with ground-truth parameter-level localization. LACUNA injects PII of synthetic individuals into predefined parameters of 1B and 7B OLMo-based models via masked continual pretraining, enabling direct evaluation of whether unlearning targets the weights responsible for knowledge storage. We use LACUNA to benchmark current SOTA unlearning methods and find that, despite strong output-level performance, existing methods are highly imprecise and susceptible to resurfacing attacks. We further show that when localization is successful, even a simple gradient-based unlearning method achieves strong erasure and robustness to resurfacing attacks, highlighting the importance of precise unlearning. We release LACUNA to complement behavioral evaluations and drive further advances in robust, localization-based unlearning.

This work studies the hidden-state inversion problem: recovering the original input token sequence of a decoder-only language model from its last-layer hidden states. Rather than treating inversion as a one-shot reconstruction, we study it as a continuous embedding-space optimisation in which a soft proxy is driven towards the leaked target without any hard-token projection during the search, and a token is committed only once, at the end of the inner loop. This design choice has two consequences which are the main focus of this paper. First, keeping the optimisation entirely in continuous space exposes a rich set of internal signals: rank trajectories of the ground-truth token, per-position loss curves, and a discrete loss measured at commit time. Second, the discrete loss allows assessing the correctness of recovery via cumulative discrete loss. We further analyse which tokens break the reconstructions and find a sharp categorical asymmetry: space-prefixed, high-frequency function words in dense regions of the embedding matrix dominate the failures, while content-bearing tokens are recovered almost perfectly. On 10-token C4 prompts the exact-match rate rises from 66.9% to 97.5% (mean similarity 0.994) as the candidate window is widened, confirming that most errors are recoverable near-misses rather than genuine ambiguities. A comparison with the released SIPIT reference situates these findings: per-step hard projection is faster, but the continuous formulation is what makes the optimisation observable and its failures detectable. The results show that last-layer hidden states of GPT-2 are as sensitive as the original text.

We audit a fixed prefix-window mean-NLL memorization probe (K=20) on a Qwen2.5-VL-7B canary testbed and report three post-hoc cases where it disagrees with full-span secret NLL or greedy exact-recall. C3 (false negative, window truncation): damage lands on hex tokens outside K=20; the probe stays flat while hit@1 drops. C4 (false positive, non-secret drift): the probe moves, but approximately 99% sits on non-secret preamble; the secret span and hit@1 are unchanged. C5 (ambiguous in-window drop): the probe falls on an undertrained baseline while full-span hex is positive and hit@1=0. Recommendation: report (i) full-span secret NLL, (ii) a span-localised decomposition, (iii) behavioural exact-recall at k>=4, and (iv) decoy probes before asserting secret-specificity. Evidence is on controlled canaries in one backbone; magnitudes are testbed-specific.

Large language models (LLMs) have increasingly moved from standalone text generation systems to agents that invoke external tools, access environments, and execute multi-step tasks. However, conventional function-calling benchmarks mainly evaluate task completion and API correctness, while privacy evaluation benchmarks typically focus on final responses or privacy judgments. Neither perspective captures purpose-bound information flow across an executed multi-tool trajectory. Motivated by this limitation in current agent evaluation, ToolPrivacyBench audits whether task-private atoms are routed only to authorized tools and downstream sinks, thereby evaluating both task completion and privacy over-disclosure during tool use. The benchmark contains 2,150 cases, including 1,150 fully synthetic privacy-sensitive business workflows and 1,000 cases adapted from existing multi-tool and function-calling benchmarks. Each case is represented by a policy knowledge base. After an agent executes against mock business backends, the evaluator compares recorded tool arguments and backend audit logs with this policy knowledge base. The evaluation covers nine widely used agents to characterize purpose-bound privacy over-disclosure. The results show that successful tool execution does not imply appropriate privacy disclosure: an agent may complete a task while transmitting unnecessary private information through intermediate tool calls. ToolPrivacyBench therefore formalizes a need-to-know disclosure boundary, under which each tool should receive only the information necessary for its stated purpose, and uses trajectory-level auditing to identify privacy over-disclosure in multi-tool workflows.

Large language model agents increasingly query databases, search document collections, call external APIs, remember past interactions, and act on a user's behalf. As they move from answering questions to operating over sensitive data, privacy becomes harder to enforce. An agent touches many data sources, runs multi-step workflows, keeps state across sessions, and acts with delegated permissions. Sensitive information can therefore leak not only through its final answer but through the queries it issues, the intermediate results it handles, the memory it writes, and the messages it exchanges with other agents. We survey the privacy of LLM agents from a data-centric view, organizing the field around the data an agent touches rather than by attack type, and we use data agent as shorthand for an LLM agent that works with data. Research on these risks is active but scattered across retrieval-augmented generation, text-to-SQL interfaces, agent memory, prompt injection, access control, and contextual privacy. This survey brings that work together: we taxonomize the data sources an agent touches, the privacy risks each source creates, and the governance mechanisms that address them; we map the benchmarks used to measure these risks and identify what is missing; and we set out the open problems. Two findings recur: among governance mechanisms only information-flow control covers both compositional and cross-session inference leakage, the two least-protected risks; and no benchmark drives an agent across its data surfaces under one privacy policy, the instrument the field most lacks. Our goal is a reference that situates the scattered literature and gives future work a common framing.

Retrieval-Augmented Generation enhances large language models by incorporating external knowledge, but deploying it in sensitive scenarios risks privacy leakage via malicious prompts. To address this, we propose a multi-agent framework that sanitizes retrieved content through semantic rewriting. By employing three specialized agents for privacy extraction, semantic analysis, and reconstruction, our approach collaboratively removes sensitive identifiers while preserving the semantic core. We evaluate the framework on the ChatDoctor and Wiki-PII datasets across six large language models. Experimental results demonstrate a significant reduction in privacy leakage under targeted attacks. For instance, we reduced targeted information exposure in LLaMA-3-8B from 144 instances in the baseline to just 1. Furthermore, we maintain strong contextual fidelity with a BLEU-1 score of 0.122, outperforming the existing SAGE method's 0.117. Finally, the framework operates as an asynchronous preprocessing module, introducing no additional latency to online inference, as all rewriting is executed as a one-time offline preprocessing step. To promote reproducibility, the source code of this work is publicly available at https://github.com/foursoils/Privacy-Preserving-RAG.

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.

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.

Large Language Models (LLMs) raise growing concerns about privacy leakage and copyright compliance. Membership inference is a key tool for assessing such risks, but existing studies mainly focus on whether specific samples or sample-based data units are used for training. We argue that LLMs exhibit a human-memory-like behavior: an LLM may not memorize a specific sample verbatim, yet it can accumulate and reveal knowledge about a real-world entity from scattered mentions. This analogy motivates us to examine whether an LLM can be interrogated like a human interviewee to reveal its exposure to entity-related information. Motivated by this question, we propose entity-level membership inference, which determines whether information related to a target entity is used in LLM training. We study this task in the practical label-only black-box setting, where only generated texts are observable. We formalize the task under clue, input, and model constraints, establish the necessary and sufficient conditions for its feasibility, and instantiate five interrogation strategies based on this formalization. The strategies use limited entity clues to construct prompts, elicit entity-related responses, and infer membership from semantic features among the generated texts. We construct entity-level datasets and adapt state-of-the-art sample-level label-only methods to the entity-level setting as baselines. Experiments on person entities show that our methods achieve AUC up to 0.97 and bring gains of 6.0%--17.5% in Balanced Accuracy over the best adapted baseline.

AI agents acting on behalf of users are constantly making decisions, and for users to trust their agents, those decisions must align with what they actually want. Privacy is an important alignment problem for agents: every message, post, or tool call an agent makes is a contextual judgment about what is appropriate to share, with whom, and under which conditions. Because such judgments depend on social expectations and norms, human judgment does not merely label privacy violations but also helps define them. While existing work relies on unreliable proxies for both training and evaluation, we place human judgment at the center of agentic privacy alignment. We introduce PrivacyAlign, a dataset of 1,350 samples with 3,516 detailed annotations from 599 unique annotators across diverse scenarios where current LLMs actually leak, and use it to ground both alignment training and automated evaluation in human privacy norms. Building on these annotations, we first show that conditioning LLM judges on human annotations and explanations for reference responses to the same prompt makes their judgments more reliable. We then introduce annotation-conditioned reward modeling, which uses these annotations to score new responses during RL, and show that small open-weight agents trained with this reward better align with human privacy norms, with strong gains on PrivacyAlign and existing privacy benchmarks for agents.

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.

Modern conversational AI systems frequently rely on user metadata to localize responses, yet the unintended regional biases introduced by this hidden context remain poorly understood. In this work, we evaluate location leakage: the phenomenon where a model generates geographic references despite receiving a geographically neutral user prompt. Across both creative writing and open-ended Q&A prompts, even state-of-the-art LLMs systematically favor region-specific outputs when exposed to location metadata, with leakage spiking by up to 793 times above baseline (e.g., from 0.04% to 31.7% for Llama 3.1-8B, and 21.3% and 8.8% for Qwen3-8B and Claude Sonnet 4.6, respectively). Our analysis further shows a novel structural conditioning effect: replacing the injected location with the placeholder "Unknown" still elevates leakage by up to 72 times above baseline, demonstrating that the user profile frame itself, independent of any geographic content, acts as a generative conditioning signal.

Split learning provides a practical paradigm for resource-constrained users to train Large Language Models (LLMs) by offloading computation-intensive layers to a server while keeping raw data local. However, existing privacy-preserving split learning methods still face a difficult trade-off among utility, privacy, efficiency, and stability. Specifically, these methods often suffer from substantial utility degradation, remain vulnerable to advanced data reconstruction attacks, incur prohibitive computational and communication overhead, or exhibit unstable performance across different tasks. In this paper, we propose MIXGUARD, a novel mixup-based privacy-preserving split learning framework for LLMs. MIXGUARD introduces token-level obfuscation, representation-level obfuscation, and adaptive gradient perturbation mechanisms, which operate jointly to preserve useful learning signals while preventing privacy leakage to the server. Technically, MIXGUARD first constructs a lightweight calibration model on a public dataset to refine the approximated target representation, and then applies this model during privacy-preserving fine-tuning on private data. We conduct extensive experiments on four classification tasks and four text generation tasks across multiple LLM families, model sizes, architectures, and fine-tuning strategies. The results show that MIXGUARD preserves model utility comparable to non-split training baselines, consistently achieves stronger privacy protection than existing split learning defense methods against state-of-the-art data reconstruction attacks, and remains robust under adaptive attack settings.

Running large language models locally is often impractical, pushing inference on sensitive text to third-party providers. Split inference partially mitigates this by keeping tokens on the client and sending only hidden representations, but these representations can still be recovered via nearest-neighbor search against the public embedding table. We propose an orthogonal obfuscation procedure in which the client multiplies embeddings by a secret orthogonal matrix before transmission. To enable correct inference under arbitrary rotations, we introduce ConjFormer, a transformer variant that is exactly $\mathrm{O}(d)$-equivariant via a lightweight normalization change (scalar RMSNorm) together with blockwise orthogonal conjugation of all linear weights. As a result, the server performs the full forward pass entirely in the rotated basis and never observes unrotated hidden states. Experiments on GPT-2 and Llama 3.2 1B models fine-tuned on PubMed show that orthogonal obfuscation eliminates direct cosine nearest-neighbor inversion and reduces token recovery from over 35% top-10 to at most 1.3%, while increasing perplexity by only 0.4% after fine-tuning. These results indicate that enforcing symmetry at the architectural level can provide a practical defense for privacy-preserving LLM inference without noise injection or heavy cryptographic machinery.

Large Language Models are increasingly trained on proprietary or sensitive data, from private healthcare and financial records to user conversations containing secrets. Ensuring the privacy of such data against extraction attacks has become a central concern. In this paper, we ask whether an attacker who can poison a portion of the training data can facilitate the leakage of a separate target record they have no access to. We answer in the affirmative and show that such leakage can be induced by a poisoning mechanism that reshapes the model's local loss landscape around the target completion. Our key insight is that poisoning to create a sharp loss minimum at the target, surrounded by elevated loss on nearby alternatives, forces the model to memorize the target as the unique low-loss solution in its neighborhood. The attack requires no architectural changes, and generalizes across centralized and federated learning settings. We demonstrate that the attack amplifies privacy leakage across language (up to 100% successful extraction), and vision-language models (up 90% successful extraction). We show that the attack is thwarted when the model is trained to be differentially private. However, we introduce a new attack that directly probes the loss landscape bypassing even differential privacy defenses.

Vision-Language-Action (VLA) models enable end-to-end robot control and have garnered widespread attention. However, the memorization of training data inherent to VLA, coupled with the high cost of robotic data acquisition, raises serious concerns regarding data privacy leakage and intellectual property infringement. Membership inference attacks (MIAs) aim to determine whether a given sample belongs to the training set. While representing a significant privacy threat, this attack remains underexplored in the context of VLA models. To bridge this gap, we propose VLALeaks, which is based on attention discrepancies in VLA models. We reveal, for the first time, the privacy vulnerabilities of VLA models. Specifically, it comprises a two-stage process: (1) membership feature extraction, and (2) attack model construction. Experimental results across multiple VLA benchmarks demonstrate that VLALeaks readily reveals membership information and achieves optimal attack AUC and TPR@1%FPR, highlighting the privacy vulnerabilities in current VLA model deployments. Our work is the first systematic study of MIAs on VLA models, aiming to provide insights for secure and trustworthy VLA models.

Large language models (LLMs) are increasingly deployed in privacy-sensitive domains, where users must balance the risk of data exposure through external APIs against the high computational cost of local deployment. Split learning has therefore emerged as a promising paradigm for LLM fine-tuning and inference under limited local resources. However, it introduces new privacy risks. Prior work primarily studies leakage of private input prompts, typically via inversion attacks on intermediate representations, while the potential for sensitive information leakage through generative response outputs remains largely unexplored. In this work, we unveil novel vulnerabilities of Split-LLM by presenting Patched Model Inversion with Dual-Sided Initialization (PIDI), a two-stage attack that simultaneously targets both private input prompts and output responses in Split-LLM settings. It combines dual-sided initialization with a patched inversion strategy to tackle long sequences, substantially outperforming prior inversion methods. To counter threats from both sides, we further propose the Adapter-based DualGuard with Mutual Information Defense (ADMI), which integrates an adapter-based local warmup strategy and mutual information regularization to provide a strong empirical privacy protection with minimal impact on task performance. Extensive experiments across diverse tasks and models demonstrate that ADMI effectively defends against PIDI and other state-of-the-art inversion attacks. Our code is publicly available at https://github.com/FLAIR-THU/VFLAIR-LLM.

A vision encoder compresses image pixels into semantic embeddings, implicitly acting as a privacy boundary by preserving semantic content while attenuating pixel-local detail required for exact text recovery. Encoder-free vision-language models (VLMs) remove this boundary by routing image patches directly into the language-model token stream, thereby exposing an architectural privacy attack surface: intermediate visual tokens become a pre-output side channel. Under a token-access adversary, decoders invert visual-token streams from two encoder-free VLMs, Gemma4 and Fuyu, recovering recognizable image structure and readable held-out access codes, whereas matched encoder-based controls localize target regions but recover no exact strings. Within-model ablations show that the operative factor is spatial sampling fidelity of the visual-token grid, especially character-direction sampling density, rather than token or value count. The leakage is not limited to exported tokens: Gemma4 layer-0 key-value cache tensors are directly invertible, placing the side channel within KV caches commonly persisted by production serving stacks for decoding efficiency. The attack survives clutter, realistic document degradation, and zero-shot transfer to public document images, and it resists value-level defenses such as additive noise and quantization. Effective mitigation must therefore reduce spatial sampling, making removal of the vision encoder a first-class privacy decision in VLM deployment.

Parameter-efficient fine-tuning of large language models (LLMs) can exhibit problematic memorization of individual training examples. Empirical privacy auditing (EPA) quantifies this risk by measuring realistic data leakage on membership inference (MI) or reconstruction attacks. A key challenge in EPA is designing ``canary'' examples that are mixed with the privacy-sensitive training data. We propose generating synthetic canaries via high-temperature sampling ($T \geq 0.8$) from LLMs, using prompts tailored to the privacy-sensitive training data. These canaries act as high-influence outliers, ensuring high identifiability and hence strong audits. Further, since the canaries are themselves non-private, they are inspectable and can be inserted with repetition without jeopardizing the privacy of the real data. An important use of models fine-tuned on privacy-sensitive data is the generation of synthetic data. This also comes with privacy risk. We introduce a powerful synthetic data audit based on fine-tuning an auxiliary model on the synthetic data. Auditing the auxiliary model for the original canaries then provides a strong estimate of the privacy leakage through the synthetic data. Finally, leveraging our strong auditing methodologies, we perform a systematic investigation into the interacting effects of model capacity and canary entropy on memorization.

Large language models (LLMs) are increasingly fine-tuned on domain-specific datasets that may contain sensitive, dataset-level properties. Recent work has shown that such dataset-level information can be effectively extracted through property inference attacks, posing a confidentiality risk. Existing defenses against these attacks primarily operate by modifying the training data distribution and hence require access to the original data and retraining the model, limiting their applicability to settings where data is unavailable or models are already deployed. In this work, we propose alignment-based defenses for mitigating property inference attacks in LLMs. Our approach reshapes the model's output distribution towards a target property ratio via post-training alignment, without modifying the training data. In particular, we adapt two widely used RLHF frameworks--Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO)--as our defenses by constructing preference pairs and defining a specific reward function respectively. Through comprehensive experiments, we show that our alignment based defenses effectively mitigate property inference attacks while maintaining a strong utility confidentiality tradeoff.

Foundation-model agents are increasingly long-lived systems that remember users across interactions, making memorization an explicit deployment-time function rather than solely a property of model weights. Existing work addresses parametric memorization or audits fixed memory configurations, but does not characterize how memory-design choices jointly shape personalization utility, extraction risk, and deletion fidelity. We study this surface as deployment-time memorization, formulating agent memory as a privacy-utility frontier measured by Personalization Recall (PR) and Adversarial Extraction Rate (AER), and sweeping three memory-design knobs: summarization aggressiveness, retrieval breadth (k), and deletion mode. We further introduce the Forgetting Residue Score (FRS) to quantify whether deleted information remains recoverable from derived memory tiers. On LongMemEval, key-fact summarization reduces canary extraction by 76% on Gemma 3 12B and 64% on GPT-4o-mini while preserving nearly all personalization recall; critically, once content is compressed away, increasing k no longer restores leakage. The same compression, however, induces a deletion-fidelity failure: raw-only deletion leaves derived summary copies recoverable in approximately 20% of instances, and only full-pipeline purge or tombstone redaction drives worst-tier residue to zero. Together, these results establish that persistent agent memory must be evaluated as a first-class memorization mechanism -- assessed by what it helps agents recall, what it makes extractable, and what it can truly erase.

Medical language models (LMs) can memorize and reproduce protected health information, but privacy evaluations often focus on recovery of training text rather than disclosure under realistic threat models. We introduce a clinically grounded framework that evaluates leakage along a graded axis of adversarial access, ranging from publicly inferable demographics to leaked note fragments. At each tier, we measure verbatim memorization of patient-specific text and semantic leakage of sensitive diagnoses. Applying the framework to an LM pretrained on 378k clinical notes, we find that routine encounter metadata (i.e. name, date of birth, provider, practice, visit date) elicits high rates of verbatim memorization across a patient's timeline and sensitive-diagnosis recovery (AUROC 0.91 for abortion, 0.81 for HIV). At the same time, exact-match memorization can overstate disclosure: 36% of memorized tokens reflect templated documentation. Our work highlights the risks of training on longitudinal clinical data, providing a practical framework for contextual privacy evaluation of medical LMs.

Tool-using large language model (LLM) agents face two distinct security failures: unauthorized external actions and exposure of sensitive plaintext inside the runtime before any final output check can intervene. Existing defenses usually protect one boundary, either the planner/runtime or the action sink, and therefore do not by themselves secure both surfaces. We present SecureClaw, a dual-boundary architecture that places authorization at the effect sink and plaintext confinement at the read boundary. Sensitive reads pass through a trusted gateway that replaces raw values with opaque handles and, in the evaluated deployment, bounded summaries as an explicit declassification interface. Writes that change external state follow a PREVIEW$\rightarrow$COMMIT protocol in which only a trusted executor may commit the exact canonical request authorized by policy. The runtime can still plan over summaries and symbolic references, but cannot directly dereference secrets or perform side effects. Across AgentDojo, AgentLeak, and Agent Security Bench (ASB), SecureClaw is the only defense we evaluate in a common harness that simultaneously retains usable task utility and achieves 0% attack success rate (ASR) on ASB, 0.64% ASR on AgentDojo, and 3.23% overall leak on AgentLeak's attacked parity lane, which measures final-output and internal-relay leakage.

Recent work has applied differential privacy (DP) to adapt large language models (LLMs) for sensitive applications, offering theoretical guarantees. However, its practical effectiveness remains unclear, partly due to LLM pretraining, where overlaps and interdependencies with adaptation data can undermine privacy despite DP efforts. To analyze this issue in practice, we investigate privacy risks under DP adaptations in LLMs using state-of-the-art attacks such as robust membership inference and canary data extraction. We benchmark these risks by systematically varying the adaptation data distribution, from exact overlaps with pretraining data, through in-distribution (IID) cases, to entirely out-of-distribution (OOD) examples. Additionally, we evaluate how different adaptation methods and different privacy regimes impact the vulnerability. Our results show that distribution shifts strongly influence privacy vulnerability: the closer the adaptation data is to the pretraining distribution, the higher the practical privacy risk at the same theoretical guarantee, even without direct data overlap. We find that parameter-efficient fine-tuning methods, such as LoRA, achieve the highest empirical privacy protection for OOD data. Our benchmark identifies key factors for achieving practical privacy in DP LLM adaptation, providing actionable insights for deploying customized models in sensitive settings. Looking forward, we propose a structured framework for holistic privacy assessment beyond adaptation privacy, to identify and evaluate risks across the full pretrain-adapt pipeline of LLMs.

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.

We present LoRA-MINT, a new methodology for Membership Inference Test (MINT) applied to recent Large Language Models (LLMs) fine-tuned for specific Natural Language Processing (NLP) tasks through Low-Rank Adaptation (LoRA). The primary goal is to assess whether individual samples were part of the training data of these adapted models, providing a useful auditing tool for the management of intellectual property and sensitive data. Our analysis explores the relationship between model perplexity and membership status, providing a systematic framework for estimating data exposure in fine-tuned LLMs. We conducted experiments on four models and three benchmark datasets, obtaining precision values in determining if given data were used for training ranging from 0.77 to 0.92, which outperform state-of-the-art baselines and demonstrate the robustness and generality of the proposed method. In general, our findings underscore the potential of LoRA-MINT as an effective and scalable framework for auditing LLMs, improving transparency, and fostering the ethical and responsible deployment of AI and NLP technologies. For the sake of concreteness and current relevance, our discussion and experiments are centered on LoRAadjusted LLMs, but note that most of the presented methodology is easily applicable for auditing training data given any other technique for adapting LLMs or, more generally, any other domain-adapted AI models.

We present the Membership Inference Test (MINT) Demo 2, a framework designed to improve transparency in machine learning training processes. MINT is a technique for experimentally determining whether specific data were used during machine learning model training. We establish the theoretical framework and propose multiple architectures for MINT depending on the amount of information known about the models that are being audited. Experimental results using a popular face recognition model, 4 state-of-the-art LLMs, and multiple, diverse, and large-scale public image and text databases achieve promising accuracy levels in the detection of training data of up to 90%. Building on these results, we introduce a comprehensive web platform1 that expands these capabilities to image and text modalities. The platform integrates a diverse technological stack, including MINT, aMINT, and gMINT, allowing users to audit a wide range of models. This demonstrator aims to promote AI transparency and provides a practical tool to foster compliance with emerging AI regulations.

Authorship attribution research has traditionally focused on long-form, expressive texts; however, interactions with large language models (LLMs) are typically brief and task-driven prompts. This raises a fundamental question: do such prompts contain a stable, author-identifiable, and distinctive signal? We introduce PromptPrint, a systematic study of prompt-based identity, the hypothesis that a user's habitual vocabulary, syntax, and discourse patterns form a learnable behavioral biometric. Using 20,680 real prompts from 1,034 users, we establish three key findings. First, lexical representations significantly outperform semantic encoders, supporting the "lexical stability hypothesis": identity is primarily encoded in surface-level word choice rather than abstract intent. Second, stylometric features exhibit a "uniqueness-consistency paradox": users are highly distinctive across the population, yet behaviorally inconsistent across contexts. Third, adversarial analysis reveals a clear vulnerability spectrum: identity signals are robust to minor lexical perturbations but degrade substantially under semantic paraphrasing. Overall, our results demonstrate strong identification performance at scale, establishing prompt-based identity as a viable behavioral biometric. This work introduces a new perspective on user modeling in LLM interactions, with important implications for security and privacy. Data and code will be released upon the acceptance of our work.

Large language models can reproduce training data, but existing memorization evaluations mostly measure whether models can be forced to do so, rather than whether they do so under ordinary use. We introduce PropMe, a propensity-aware framework for memorization evaluation that contrasts prefix-based capability attacks with non-adversarial evaluations. We propose a metric transformation that, applied to existing functions, allows to create propensity metrics. We further introduce SimpleTrace, a lightweight tracing pipeline built on infini-gram that deterministically attributes model generations to large-scale training corpora and computes verbatim, near-verbatim, and propensity-transformed memorization metrics. Evaluating two fully-open models: Comma and DFM Decoder on two datasets: Common Pile and Dynaword in two languages, we find a consistent gap between capability and propensity: prefix attacks elicit substantially stronger memorization signals than generic or dataset-specific prompts, while propensity scores remain low overall. Thus, the models can reveal training data when directly elicited, but rarely do so in more common non-adversarial settings. We also find that DFM Decoder, which is continually pre-trained from Comma, exhibits reduced memorization and memorization propensity for Common Pile, confirming that memorization capability can decrease when later training emphasizes partially different data. Our results suggest, and we encourage, that memorization audits should report both worst-case extractability and ordinary leakage propensity in order to have a more comprehensive view of this phenomenon.

Long-term memory enables language model agents to support personalized interactions, but it remains unclear when available memories warrant integration into responses. Existing memory evaluations emphasize retrieval accuracy and downstream task utility, while overlooking whether retrieved sensitive memory content is warranted in the current turn. We introduce RBI-Eval, a controlled measurement study built around a probe set that compares model behavior with and without access to sensitive memory under identical benign prompts. We evaluate four base LLMs against a matched no-memory reference across four memory-access settings: full-context exposure and three retrieval systems. Our results reveal substantial behavioral divergence. With memory available, the separation score for sensitive-memory integration decreases by 8.9%--26.6% relative to the matched no-memory reference for GPT-5.4-mini, but by 51.1%--82.9% for Claude-Sonnet-4.6, DeepSeek-V4-Flash, and Qwen3.5-9B. Control experiments on DeepSeek and GPT-5.4-mini show this effect is specific to sensitive content, rather than general personalization. Retrieval systems reduce exposure but do not eliminate integration once sensitive memory reaches the generator. These findings suggest safe personalization requires memory-aware decisions at both retrieval and generation time.