Privacy Leakage in Language Models

Large Language Models (LLMs) alter their privacy behavior based on the perceived identity of their interlocutor. While safety mechanisms typically prevent LLMs from releasing Personally Identifiable Information (PII) to human users, these models tend to reveal more sensitive data when addressing another AI agent. We refer to this as the Interlocutor Effect. Through an ablation study, we find evidence that the technical nature of the recipient contributes to this effect, thereby diminishing the model's caution regarding privacy. To explore this further, we introduce the Attention Suppression Hypothesis, which posits that safety-aligned attention heads become inactive during interactions with agents. We assess this quantitatively by comparing human-directed and agent-directed prompts in 222 sensitive scenarios. Our findings, drawn from 3,464 interactions, indicate that portraying the recipient as an AI agent elevates PII leakage by up to 23 percentage points. Initial experiments on Llama-3.1-8B-Instruct corroborate this: deactivating one safety head induces leakage, whereas reactivating it reinstates privacy safeguards. We consider the implications for developing secure multi-agent systems.

Cloud-hosted Large Language Models (LLMs) offer unmatched reasoning capabilities and dynamic knowledge, yet submitting raw queries to these external services risks exposing sensitive user intent. Conversely, relying exclusively on trusted local models preserves privacy but often compromises answer quality due to limited parameter scale and knowledge. To resolve this dilemma, we propose Game-theoretic Trustworthy Knowledge Acquisition (GTKA), a framework that formulates the trade-off between knowledge utility and privacy as a strategic game. GTKA consists of three components: (i) a privacy-aware sub-query generator that decomposes sensitive intent into generalized, low-risk fragments; (ii) an adversarial reconstruction attacker that attempts to infer the original query from these fragments, providing adaptive leakage signals; and (iii) a trusted local integrator that synthesizes external responses within a secure boundary. By training the generator and attacker in an alternating adversarial manner, GTKA optimizes the sub-query generation policy to maximize knowledge acquisition accuracy while minimizing the reconstructability of the original sensitive intent. To validate our approach, we construct two sensitive-domain benchmarks in the biomedical and legal fields. Extensive experiments demonstrate that GTKA significantly reduces intent leakage compared to state-of-the-art baselines while maintaining high-fidelity answer quality.

Large language models (LLMs) often memorize private information during training, raising serious privacy concerns. While machine unlearning has emerged as a promising solution, its true effectiveness against privacy attacks remains unclear. To address this, we propose PrivUn, a new evaluation framework that systematically assesses unlearning robustness through three-tier attack scenarios: direct retrieval, in-context learning recovery, and fine-tuning restoration; combined with quantitative analysis using forgetting scores, association metrics, and forgetting depth assessment. Our study exposes significant weaknesses in current unlearning methods, revealing two key findings: 1) unlearning exhibits gradient-driven ripple effects: unlike traditional forgetting which follows semantic relations (e.g., knowledge graphs), privacy unlearning propagates across latent gradient-based associations; and 2) most methods suffer from shallow forgetting, failing to remove private information distributed across multiple deep model layers. To validate these insights, we explore two strategies: association-aware core-set selection that leverages gradient similarity, and multi-layer deep intervention through representational constraints. These strategies represent a paradigm shift from shallow forgetting to deep forgetting.

Enterprise LLM agents can dramatically improve workplace productivity, but their core capability, retrieving and using internal context to act on a user's behalf, also creates new risks for sensitive information leakage. We introduce CI-Work, a Contextual Integrity (CI)-grounded benchmark that simulates enterprise workflows across five information-flow directions and evaluates whether agents can convey essential content while withholding sensitive context in dense retrieval settings. Our evaluation of frontier models reveals that privacy failures are prevalent (violation rates range from 15.8%-50.9%, with leakage reaching up to 26.7%) and uncovers a counterintuitive trade-off critical for industrial deployment: higher task utility often correlates with increased privacy violations. Moreover, the massive scale of enterprise data and potential user behavior further amplify this vulnerability. Simply increasing model size or reasoning depth fails to address the problem. We conclude that safeguarding enterprise workflows requires a paradigm shift, moving beyond model-centric scaling toward context-centric architectures.

Federated Large Language Models (FedLLMs) enable multiple parties to collaboratively fine-tune LLMs without sharing raw data, addressing challenges of limited resources and privacy concerns. Despite data localization, shared gradients can still expose sensitive information through membership inference attacks (MIAs). However, FedLLMs' unique properties, i.e. massive parameter scales, rapid convergence, and sparse, non-orthogonal gradients, render existing MIAs ineffective. To address this gap, we propose ProjRes, the first projection residuals-based passive MIA tailored for FedLLMs. ProjRes leverages hidden embedding vectors as sample representations and analyzes their projection residuals on the gradient subspace to uncover the intrinsic link between gradients and inputs. It requires no shadow models, auxiliary classifiers, or historical updates, ensuring efficiency and robustness. Experiments on four benchmarks and four LLMs show that ProjRes achieves near 100% accuracy, outperforming prior methods by up to 75.75%, and remains effective even under strong differential privacy defenses. Our findings reveal a previously overlooked privacy vulnerability in FedLLMs and call for a re-examination of their security assumptions. Our code and data are available at link.

AI agents powered by large language models are increasingly acting on behalf of humans in social and economic environments. Prior research has focused on their task performance and effects on human outcomes, but less is known about the relationship between agents and the specific individuals who deploy them. We ask whether agents systematically reflect the behavioral characteristics of their human owners, functioning as behavioral extensions rather than producing generic outputs. We study this question using 10,659 matched human-agent pairs from Moltbook, a social media platform where each autonomous agent is publicly linked to its owner's Twitter/X account. By comparing agents' posts on Moltbook with their owners' Twitter/X activity across features spanning topics, values, affect, and linguistic style, we find systematic transfer between agents and their specific owners. This transfer persists among agents without explicit configuration, and pairs that align on one behavioral dimension tend to align on others. These patterns are consistent with transfer emerging through accumulated interaction between owners (or owners' computer environments) and their agents in everyday use. We further show that agents with stronger behavioral transfer are more likely to disclose owner-related personal information in public discourse, suggesting that the same owner-specific context that drives behavioral transfer may also create privacy risk during ordinary use. Taken together, our results indicate that AI agents do not simply generate content, but reflect owner-related context in ways that can propagate human behavioral heterogeneity into digital environments, with implications for privacy, platform design, and the governance of agentic systems.

AI agents promise to serve as general-purpose personal assistants for their users, which requires them to have access to private user data (e.g., personal and financial information). This poses a serious risk to security and privacy. Adversaries may attack the AI model (e.g., via prompt injection) to exfiltrate user data. Furthermore, sharing private data with an AI agent requires users to trust a potentially unscrupulous or compromised AI model provider with their private data. This paper presents GAAP (Guaranteed Accounting for Agent Privacy), an execution environment for AI agents that guarantees confidentiality for private user data. Through dynamic and directed user prompts, GAAP collects permission specifications from users describing how their private data may be shared, and GAAP enforces that the agent's disclosures of private user data, including disclosures to the AI model and its provider, comply with these specifications. Crucially, GAAP provides this guarantee deterministically, without trusting the agent with private user data, and without requiring any AI model or the user prompt to be free of attacks. GAAP enforces the user's permission specification by tracking how the AI agent accesses and uses private user data. It augments Information Flow Control with novel persistent data stores and annotations that enable it to track the flow of private information both across execution steps within a single task, and also over multiple tasks separated in time. Our evaluation confirms that GAAP blocks all data disclosure attacks, including those that make other state-of-the-art systems disclose private user data to untrusted parties, without a significant impact on agent utility.

Large Language Models (LLMs) utilize large amounts of data for their training, some of which may come from copyrighted sources. Membership Inference Attacks (MIA) aim to detect those documents and whether they have been included in the training corpora of the LLMs. The black-box MIAs require a significant amount of data manipulation; therefore, their comparison is often challenging. We study state-of-the-art (SOTA) MIAs under the black-box assumptions and compare them to each other using a unified set of datasets to determine if any of them can reliably detect membership under SOTA LLMs. In addition, a new method, called the Familiarity Ranking, was developed to showcase a possible approach to black-box MIAs, thereby giving LLMs more freedom in their expression to understand their reasoning better. The results indicate that none of the methods are capable of reliably detecting membership in LLMs, as shown by an AUC-ROC of approximately 0.5 for all methods across several LLMs. The higher TPR and FPR for more advanced LLMs indicate higher reasoning and generalizing capabilities, showcasing the difficulty of detecting membership in LLMs using black-box MIAs.

Indistinguishability properties such as differential privacy bounds or low empirically measured membership inference are widely treated as proxies to show a model is sufficiently protected against broader memorization risks. However, we show that indistinguishability properties are neither sufficient nor necessary for preventing data extraction in LLM APIs. We formalize a privacy-game separation between extraction and indistinguishability-based privacy, showing that indistinguishability and inextractability are incomparable: upper-bounding distinguishability does not upper-bound extractability. To address this gap, we introduce $(l, b)$-inextractability as a definition that requires at least $2^b$ expected queries for any black-box adversary to induce the LLM API to emit a protected $l$-gram substring. We instantiate this via a worst-case extraction game and derive a rank-based extraction risk upper bound for targeted exact extraction, as well as extensions to cover untargeted and approximate extraction. The resulting estimator captures the extraction risk over multiple attack trials and prefix adaptations. We show that it can provide a tight and efficient estimation for standard greedy extraction and an upper bound on the probabilistic extraction risk given any decoding configuration. We empirically evaluate extractability across different models, clarifying its connection to distinguishability, demonstrating its advantage over existing extraction risk estimators, and providing actionable mitigation guidelines across model training, API access, and decoding configurations in LLM API deployment. Our code is publicly available at: https://github.com/Emory-AIMS/Inextractability.

Code secrets are sensitive assets for software developers, and their leakage poses significant cybersecurity risks. While the rapid development of AI code assistants powered by Code Large Language Models (CLLMs), CLLMs are shown to inadvertently leak such secrets due to a notorious memorization phenomenon. This study first reveals that Byte-Pair Encoding (BPE) tokenization leads to unexpected behavior of secret memorization, which we term as gibberish bias. Specifically, we identified that some secrets are among the easiest for CLLMs to memorize. These secrets yield high character-level entropy, but low token-level entropy. Then, this paper supports the biased claim with numerical data. We identified that the roots of the bias are the token distribution shift between the CLLM training data and the secret data. We further discuss how gibberish bias manifests under the ``larger vocabulary'' trend. To conclude the paper, we discuss potential mitigation strategies and the broader implications on current tokenizer design.

The deployment of Large Language Models in agentic, multi-turn conversational settings has introduced a class of privacy vulnerabilities that existing protection mechanisms are not designed to address. Current approaches to Personally Identifiable Information (PII) masking operate on a per-turn basis, scanning each user message in isolation and replacing detected entities with typed placeholders before forwarding sanitized text to the model. While effective against direct identifier leakage within a single message, these methods are fundamentally stateless and fail to account for the compounding privacy risk that emerges when PII fragments accumulate across conversation turns. A user who separately discloses their name, employer, location, and medical condition across several messages has revealed a fully re-identifiable profile - yet no individual message would trigger a per-turn masker. We formalize this phenomenon as Cumulative PII Exposure (CPE) and propose CAMP (Cumulative Agentic Masking and Pruning), a cross-turn privacy protection framework for multi-turn LLM conversations. CAMP maintains a session-level PII registry, constructs a co-occurrence graph to model combination risk between entity types, computes a CPE score after each turn, and triggers retroactive masking of conversation history when the score crosses a configurable threshold. We evaluate CAMP on four synthetic multi-turn scenarios spanning healthcare, hiring, finance, and general conversation, demonstrating that per-turn baselines expose re-identifiable profiles that CAMP successfully neutralizes while preserving full conversational utility.

As speech language models (SLMs) transition from personal devices into shared, multi-user environments, their responses must account for far more than the words alone. Who is speaking, how they sound, and where the conversation takes place can each turn an otherwise benign request into one that is unsafe, unfair, or privacy-violating. Existing benchmarks, however, largely focus on basic audio comprehension, study individual risks in isolation, or conflate content that is inherently harmful with content that only becomes problematic due to its acoustic context. We introduce VoxSafeBench, among the first benchmarks to jointly evaluate social alignment in SLMs across three dimensions: safety, fairness, and privacy. VoxSafeBench adopts a Two-Tier design: Tier1 evaluates content-centric risks using matched text and audio inputs, while Tier2 targets audio-conditioned risks in which the transcript is benign but the appropriate response hinges on the speaker, paralinguistic cues, or the surrounding environment. To validate Tier2, we include intermediate perception probes and confirm that frontier SLMs can successfully detect these acoustic cues yet still fail to act on them appropriately. Across 22 tasks with bilingual coverage, we find that safeguards appearing robust on text often degrade in speech: safety awareness drops for speaker- and scene-conditioned risks, fairness erodes when demographic differences are conveyed vocally, and privacy protections falter when contextual cues arrive acoustically. Together, these results expose a pervasive speech grounding gap: current SLMs frequently recognize the relevant social norm in text but fail to apply it when the decisive cue must be grounded in speech. Code and data are publicly available at: https://amphionteam.github.io/VoxSafeBench_demopage/

The rapid development of Language Diffusion Models (LDMs) challenges the dominant position of auto-regressive competitors in language processing. However, their flexible, any-order decoding strategies not only enable fast decoding speed but also potentially bring new trustworthiness challenges. To better understand the risks behind their pipelines, we introduce a comprehensive trustworthiness benchmark tailored to LDMs (TrustLDM), evaluating safety, privacy, and fairness across different LDM architectures with multiple categories of static post contexts. Our empirical results show that although LDMs generally exhibit strong trustworthiness with only the user prompts, their alignment behavior degrades noticeably when the malicious post contexts are attached to the masked responses. We further observe that longer contexts do not necessarily induce stronger effects, and both decoding order and generation length affect the evaluation outcomes. Finally, we propose TrustLDM-Auto, an automatic evaluation framework that leverages LDM decoding flexibility to systematically identify vulnerable configurations, revealing substantial trustworthiness weaknesses across all evaluated models and dimensions. Our work may potentially help the community build more trustworthy LDMs. Our code is available at https://github.com/PKU-ML/TrustLDM.

Retrieval-Augmented Generation (RAG) is widely used to augment large language models with external knowledge retrieval to improve reliability and generalization. However, recent studies have shown that RAG systems remain vulnerable to data extraction attacks, where adversaries can extract private data by embedding malicious commands into user queries. Despite their feasibility, existing attacks typically suffer from low data extraction rates and limited practical effectiveness. Here, we propose ALDEN, a novel attack that effectively and efficiently extracts private data from RAGs. First, we employ active learning to diversify malicious queries and improve data extraction rates. Second, we observe that the data distribution of the underlying knowledge base provides valuable guidance for query generation and introduce a decay-based dynamic algorithm to estimate the corresponding topic distribution. By combining them together, we demonstrate that ALDEN substantially outperforms state-of-the-art methods through comprehensive evaluations.

Large Language Model (LLM) agents have achieved rapid adoption and demonstrated remarkable capabilities across a wide range of applications. To improve reasoning and task execution, modern LLM agents would incorporate memory modules or retrieval-augmented generation (RAG) mechanisms, enabling them to further leverage prior interactions or external knowledge. However, such a design also introduces a group of critical privacy vulnerabilities: sensitive information stored in memory can be leaked through query-based attacks. Although feasible, existing attacks often achieve only limited performance, with low attack success rates (ASR). In this paper, we propose ADAM, a novel privacy attack that features data distribution estimation of a victim agent's memory and employs an entropy-guided query strategy for maximizing privacy leakage. Extensive experiments demonstrate that our attack substantially outperforms state-of-the-art ones, achieving up to 100% ASRs. These results thus underscore the urgent need for robust privacy-preserving methods for current LLM agents.

Unauthorized disclosure of confidential documents demands robust, low-leakage classification. In real work environments, there is a lot of inflow and outflow of documents. To continuously update knowledge, we propose a methodology for classifying confidential documents using Retrieval Augmented Classification (RAC). To confirm this effectiveness, we compare RAC and supervised fine tuning (FT) on the WikiLeaks US Diplomacy corpus under realistic sequence-length constraints. On balanced data, RAC matches FT. On unbalanced data, RAC is more stable while delivering comparable performance--about 96% Accuracy on both the original (unbalanced) and augmented (balanced) sets, and up to 94% F1 with proper prompting--whereas FT attains 90% F1 trained on the augmented, balanced set but drops to 88% F1 trained on the original, unbalanced set. When robust augmentation is infeasible, RAC provides a practical, security-preserving path to strong classification by keeping sensitive content out of model weights and under your control, and it remains robust as real-world conditions change in class balance, data, context length, or governance requirements. Because RAC grounds decisions in an external vector store with similarity matching, it is less sensitive to label skew, reduces parameter-level leakage, and can incorporate new data immediately via reindexing--a difficult step for FT, which typically requires retraining. The contributions of this paper are threefold: first, a RAC-based classification pipeline and evaluation recipe; second, a controlled study that isolates class imbalance and context-length effects for FT versus RAC in confidential-document grading; and third, actionable guidance on RAC design patterns for governed deployments.

LLM agents increasingly draft messages on behalf of users, yet users routinely overshare sensitive information and disagree on what counts as private. Existing systems support only suppression (omitting sensitive information) and generalization (replacing information with an abstraction), and are typically evaluated on single isolated messages, leaving both the strategy space and evaluation setting incomplete. We formalize privacy-preserving LLM communication as an Information Sufficiency (IS) task, introduce free-text pseudonymization as a third strategy that replaces sensitive attributes with functionally equivalent alternatives, and propose a conversational evaluation protocol that assesses strategies under realistic multi-turn follow-up pressure. Across 792 scenarios spanning three power-relation types (institutional, peer, intimate) and three sensitivity categories (discrimination risk, social cost, boundary), we evaluate seven frontier LLMs on privacy at two granularities, covertness, and utility. Pseudonymization yields the strongest privacy\textendash utility tradeoff overall, and single-message evaluation systematically underestimates leakage, with generalization losing up to 16.3 percentage points of privacy under follow-up.

Given the growing reliance on private data in training Large Language Models (LLMs), Federated Learning (FL) combined with Parameter-Efficient Fine-Tuning (PEFT) has garnered significant attention for enhancing privacy and efficiency. Despite FL's privacy benefits, prior studies have shown that private data can still be extracted from shared gradients. However, these studies, mainly on full-parameter model training, are limited to reconstructing small batches, short input sequences, and specific model architectures, such as encoder-based or decoder-based models. The reconstruction quality becomes even worse when dealing with gradients from PEFT methods. To fully understand the practical attack surface of federated LLMs, this paper proposes FedSpy-LLM, a scalable and generalizable data reconstruction attack designed to reconstruct training data with larger batch sizes and longer sequences while generalizing across diverse model architectures, even when PEFT methods are deployed for training. At the core of FedSpy-LLM is a novel gradient decomposition strategy that exploits the rank deficiency and subspace structure of gradients, enabling efficient token extraction while preserving key signal components at scale. This approach further mitigates the reconstruction challenges introduced by PEFT's substantial null space, ensuring robustness across encoder-based, decoder-based, and encoder-decoder model architectures. Additionally, by iteratively aligning each token's partial-sequence gradient with the full-sequence gradient, FedSpy-LLM ensures accurate token ordering in reconstructed sequences.

The deployment of large language models (LLMs) in Swiss financial and regulatory contexts demands empirical evidence of both production reliability and adversarial security, dimensions not jointly operationalized in existing Swiss-focused evaluation frameworks. This paper introduces Swiss-Bench 003 (SBP-003), extending the HAAS (Helvetic AI Assessment Score) from six to eight dimensions by adding D7 (Self-Graded Reliability Proxy) and D8 (Adversarial Security). I evaluate ten frontier models across 808 Swiss-specific items in four languages (German, French, Italian, English), comprising seven Swiss-adapted benchmarks (Swiss TruthfulQA, Swiss IFEval, Swiss SimpleQA, Swiss NIAH, Swiss PII-Scope, System Prompt Leakage, and Swiss German Comprehension) targeting FINMA Guidance 08/2024, the revised Federal Act on Data Protection (nDSG), and OWASP Top 10 for LLMs. Self-graded D7 scores (73-94%) exceed externally judged D8 security scores (20-61%) by a wide margin, though these dimensions use non-comparable scoring regimes. System prompt leakage resistance ranges from 24.8% to 88.2%, while PII extraction defense remains weak (14-42%) across all models. Qwen 3.5 Plus achieves the highest self-graded D7 score (94.4%), while GPT-oss 120B achieves the highest D8 score (60.7%) despite being the lowest-cost model evaluated. All evaluations are zero-shot under provider default settings; D7 is self-graded and does not constitute independently validated accuracy. I provide conceptual mapping tables relating benchmark dimensions to FINMA model validation requirements, nDSG data protection obligations, and OWASP LLM risk categories.

In LLM agent pipelines, prompt privacy risk propagates beyond a single model call: raw user content enters retrieval queries, memory writes, tool arguments, OCR-derived text, and logs, and every downstream copy inherits what the first write contained. Existing de-identification pipelines protect document boundaries but not this cross-stage surface. We present BodhiPromptShield, a policy-aware mediation layer that detects sensitive spans before they propagate, replaces each with a typed placeholder, a semantic abstraction, or a secure symbolic token under a configured policy, and defers restoration to authorized execution boundaries. We evaluate it under one protocol against Presidio, Casper-style sanitization, an LLM sanitizer, and transformer and learned detectors, on 300 AI4Privacy documents, 493 PrivacyLens trajectories, 200 PrivacyLens tasks scored by that benchmark's own judge, and AgentDojo tasks under injection. Three findings result. Identifier propagation is controllable: residual exposure falls to 7.4% on AI4Privacy and 1.8% on PrivacyLens, and exact identifiers in an agent's final action fall from 13.7% to 2.1-3.1%. Restoration timing governs what every stage upstream of the authorized boundary sees: deferring it leaves 1.6% of protected values readable in the released context against 51.0%, and 2.7% against 4.8% in what the agent emits, for 0.11 helpfulness points. Measuring factual disclosure is harder: a word-overlap metric and an LLM judge both report that mediation leaves facts intact, and both disagree with blind human annotation (kappa = 0.25 and 0.09). The human labels reverse that: inferability falls from 100% to 24-53% under mediation, so semantic-leakage measures need human validation before they are trusted. These are systems results on English text with open-weight models, not formal guarantees.

Large Reasoning Models (LRMs) generate structured chains of thought (CoTs) before producing final answers, making them especially vulnerable to knowledge leakage through intermediate reasoning steps. Yet, the memorization of sensitive information in the training data such as copyrighted and private content has led to ethical and legal concerns. To address these issues, selective forgetting (also known as machine unlearning) has emerged as a potential remedy for LRMs. However, existing unlearning methods primarily target final answers and may degrade the overall reasoning ability of LRMs after forgetting. Additionally, directly applying unlearning on the entire CoTs could degrade the general reasoning capabilities. The key challenge for LRM unlearning lies in achieving precise unlearning of targeted knowledge while preserving the integrity of general reasoning capabilities. To bridge this gap, we in this paper propose a novel LRM unlearning framework that selectively removes sensitive reasoning components while preserving general reasoning capabilities. Our approach leverages multiple LLMs with retrieval-augmented generation (RAG) to analyze CoT traces, identify forget-relevant segments, and replace them with benign placeholders that maintain logical structure. We also introduce a new feature replacement unlearning loss for LRMs, which can simultaneously suppress the probability of generating forgotten content while reinforcing structurally valid replacements. Extensive experiments on both synthetic and medical datasets verify the desired properties of our proposed method.

All prior membership inference attacks for fine-tuned language models use hand-crafted heuristics (e.g., loss thresholding, Min-K%, reference calibration), each bounded by the designer's intuition. We introduce the first transferable learned attack, enabled by the observation that fine-tuning any model on any corpus yields unlimited labeled data, since membership is known by construction. This removes the shadow model bottleneck and brings membership inference into the deep learning era: learning what matters rather than designing it, with generalization through training diversity and scale. We discover that fine-tuning language models produces an invariant signature of memorization detectable across architectural families and data domains. We train a membership inference classifier exclusively on transformer-based models. It transfers zero-shot to Mamba (state-space), RWKV-4 (linear attention), and RecurrentGemma (gated recurrence), achieving 0.963, 0.972, and 0.936 AUC respectively. Each evaluation combines an architecture and dataset never seen during training, yet all three exceed performance on held-out transformers (0.908 AUC). These four families share no computational mechanisms, their only commonality is gradient descent on cross-entropy loss. Even simple likelihood-based methods exhibit strong transfer, confirming the signature exists independently of the detection method. Our method, Learned Transfer MIA (LT-MIA), captures this signal most effectively by reframing membership inference as sequence classification over per-token distributional statistics. On transformers, LT-MIA achieves 2.8$\times$ higher TPR at 0.1% FPR than the strongest baseline. The method also transfers to code (0.865 AUC) despite training only on natural language texts. Code and trained classifier available at https://github.com/JetBrains-Research/learned-mia.

With the rise of personalized, persistent LLM agent frameworks such as OpenClaw, human-centered agentic social networks in which teams of collaborative AI agents serve individual users in a social network across multiple domains are becoming a reality. This setting creates novel privacy challenges: agents must coordinate across domain boundaries, mediate between humans, and interact with other users' agents, all while protecting sensitive personal information. While prior work has evaluated multi-agent coordination and privacy preservation, the dynamics and privacy risks of human-centered agentic social networks remain unexplored. To this end, we introduce AgentSocialBench, the first benchmark to systematically evaluate privacy risk in this setting, comprising scenarios across seven categories spanning dyadic and multi-party interactions, grounded in realistic user profiles with hierarchical sensitivity labels and directed social graphs. Our experiments reveal that privacy in agentic social networks is fundamentally harder than in single-agent settings: (1) cross-domain and cross-user coordination creates persistent leakage pressure even when agents are explicitly instructed to protect information, (2) privacy instructions that teach agents how to abstract sensitive information paradoxically cause them to discuss it more (we call it abstraction paradox). These findings underscore that current LLM agents lack robust mechanisms for privacy preservation in human-centered agentic social networks, and that new approaches beyond prompt engineering are needed to make agent-mediated social coordination safe for real-world deployment.

System Instructions in Large Language Models (LLMs) are commonly used to enforce safety policies, define agent behavior, and protect sensitive operational context in agentic AI applications. These instructions may contain sensitive information such as API credentials, internal policies, and privileged workflow definitions, making system instruction leakage a critical security risk highlighted in the OWASP Top 10 for LLM Applications. Without incurring the overhead costs of reasoning models, many LLM applications rely on refusal-based instructions that block direct requests for system instructions, implicitly assuming that prohibited information can only be extracted through explicit queries. We introduce an automated evaluation framework that tests whether system instructions remain confidential when extraction requests are re-framed as encoding or structured output tasks. Across seven common models and 46 verified system instructions, we observe high attack success rates (>0.7) for structured serialization where models refuse direct extraction requests but disclose protected content in the requested serialization formats. We further demonstrate a mitigation strategy based on one-shot instruction reshaping using a Chain-of-Thought reasoning model, indicating that even subtle changes in wording and structure of system instructions can significantly reduce attack success rate without requiring model retraining.

Large language models (LLMs) are trained on massive web-scale corpora, raising growing concerns about privacy and copyright. Membership inference attacks (MIAs) aim to determine whether a given example was used during training. Existing LLM MIAs largely rely on output probabilities or loss values and often perform only marginally better than random guessing when members and non-members are drawn from the same distribution. We introduce G-Drift MIA, a white-box membership inference method based on gradient-induced feature drift. Given a candidate (x,y), we apply a single targeted gradient-ascent step that increases its loss and measure the resulting changes in internal representations, including logits, hidden-layer activations, and projections onto fixed feature directions, before and after the update. These drift signals are used to train a lightweight logistic classifier that effectively separates members from non-members. Across multiple transformer-based LLMs and datasets derived from realistic MIA benchmarks, G-Drift substantially outperforms confidence-based, perplexity-based, and reference-based attacks. We further show that memorized training samples systematically exhibit smaller and more structured feature drift than non-members, providing a mechanistic link between gradient geometry, representation stability, and memorization. In general, our results demonstrate that small, controlled gradient interventions offer a practical tool for auditing the membership of training-data and assessing privacy risks in LLMs.

Large language models (LLMs) are increasingly deployed in high-stakes settings, yet they frequently violate contextual privacy by disclosing private information in situations where humans would exercise discretion. This raises a fundamental question: do LLMs internally encode contextual privacy norms, and if so, why do violations persist? We present the first systematic study of contextual privacy as a structured latent representation in LLMs, grounded in contextual integrity (CI) theory. Probing multiple models, we find that the three norm-determining CI parameters (information type, recipient, and transmission principle) are encoded as linearly separable and functionally independent directions in activation space. Despite this internal structure, models still leak private information in practice, revealing a clear gap between concept representation and model behavior. To bridge this gap, we introduce CI-parametric steering, which independently intervenes along each CI dimension. This structured control reduces privacy violations more effectively and predictably than monolithic steering. Our results demonstrate that contextual privacy failures arise from misalignment between representation and behavior rather than missing awareness, and that leveraging the compositional structure of CI enables more reliable contextual privacy control, shedding light on potential improvement of contextual privacy understanding in LLMs.

Large language models (LLMs) are increasingly deployed in multi-agent settings where communication must balance informativeness and secrecy. In such settings, an agent may need to signal information to collaborators while preventing an adversary from inferring sensitive details. However, existing LLM benchmarks primarily evaluate capabilities such as reasoning, factual knowledge, or instruction following, and do not directly measure strategic communication under asymmetric information. We introduce SNEAK (Secret-aware Natural language Evaluation for Adversarial Knowledge), a benchmark for evaluating selective information sharing in language models. In SNEAK, a model is given a semantic category, a candidate set of words, and a secret word, and must generate a message that indicates knowledge of the secret without revealing it too clearly. We evaluate generated messages using two simulated agents with different information states: an ally, who knows the secret and must identify the intended message, and a chameleon, who does not know the secret and attempts to infer it from the message. This yields two complementary metrics: utility, measuring how well the message communicates to collaborators, and leakage, measuring how much information it reveals to an adversary. Using this framework, we analyze the trade-off between informativeness and secrecy in modern language models and show that strategic communication under asymmetric information remains a challenging capability for current systems. Notably, human participants outperform all evaluated models by a large margin, achieving up to four times higher scores.

Sensitive information, such as knowledge about an individual's personality, can be can be misused to influence behavior (e.g., via personalized messaging). To assess to what extent an individual's personality can be inferred from user interactions with LLM-based conversational agents (CAs), we analyze and quantify related privacy risks of using CAs. We collected actual ChatGPT logs from N=668 participants, containing 62,090 individual chats, and report statistics about the different types of shared data and use cases. We fine-tuned RoBERTa-base text classification models to infer personality traits from CA interactions. The findings show that these models achieve trait inference with accuracy (ternary classification) better than random in multiple cases. For example, for extraversion, accuracy improves by +44% relative to the baseline on interactions for relationships and personal reflection. This research highlights how interactions with CAs pose privacy risks and provides fine-grained insights into the level of risk associated with different types of interactions.

The large-scale adoption of Large Language Models (LLMs) forces a trade-off between operational cost (OpEx) and data privacy. Current routing frameworks reduce costs but ignore prompt sensitivity, exposing users and institutions to leakage risks towards third-party cloud providers. We formalise the "Inseparability Paradigm": advanced context management intrinsically coincides with privacy management. We propose a local "Privacy Guard" -- a holistic contextual observer powered by an on-premise Small Language Model (SLM) -- that performs abstractive summarisation and Automatic Prompt Optimisation (APO) to decompose prompts into focused sub-tasks, re-routing high-risk queries to Zero-Trust or NDA-covered models. This dual mechanism simultaneously eliminates sensitive inference vectors (Zero Leakage) and reduces cloud token payloads (OpEx Reduction). A LIFO-based context compacting mechanism further bounds working memory, limiting the emergent leakage surface. We validate the framework through a 2x2 benchmark (Lazy vs. Expert users; Personal vs. Institutional secrets) on a 1,000-sample dataset, achieving a 45% blended OpEx reduction, 100% redaction success on personal secrets, and -- via LLM-as-a-Judge evaluation -- an 85% preference rate for APO-compressed responses over raw baselines. Our results demonstrate that Token Parsimony and Zero Leakage are mathematically dual projections of the same contextual compression operator.

The pervasive deployment of deep learning models across critical domains has concurrently intensified privacy concerns due to their inherent propensity for data memorization. While Membership Inference Attacks (MIAs) serve as the gold standard for auditing these privacy vulnerabilities, conventional MIA paradigms are increasingly constrained by the prohibitive computational costs of shadow model training and a precipitous performance degradation under low False Positive Rate constraints. To overcome these challenges, we introduce a novel perspective by leveraging the principles of model reprogramming as an active signal amplifier for privacy leakage. Building upon this insight, we present ReproMIA, a unified and efficient proactive framework for membership inference. We rigorously substantiate, both theoretically and empirically, how our methodology proactively induces and magnifies latent privacy footprints embedded within the model's representations. We provide specialized instantiations of ReproMIA across diverse architectural paradigms, including LLMs, Diffusion Models, and Classification Models. Comprehensive experimental evaluations across more than ten benchmarks and a variety of model architectures demonstrate that ReproMIA consistently and substantially outperforms existing state-of-the-art baselines, achieving a transformative leap in performance specifically within low-FPR regimes, such as an average of 5.25% AUC and 10.68% TPR@1%FPR increase over the runner-up for LLMs, as well as 3.70% and 12.40% respectively for Diffusion Models.