privacy compliance teams

Abstract: Prior research suggests that some dimensions of AI empathy may increase users' disclosure to chatbots, posing growing privacy risks as chatbots become more capable of empathic interaction. Yet this work often conceptualizes AI empathy narrowly, leaving users' own definitions and their broader privacy experiences unexplored. We interviewed 21 AI Chatbot users about what AI empathy meant to them and how it shaped their privacy attitudes, behaviors, and needs. Participants characterized empathy through five components--understanding, resonating, caring, expressing, and helping. Most participants viewed resonating and caring as essential components of human empathy but absent from AI empathy. Yet AI's contextual, in-depth understanding and helpfulness still encouraged highly sensitive disclosure. This highlights that AI empathy can endanger users' privacy without necessarily making AI seem human. Participants also desired varied forms and levels of empathy, each associated with distinct privacy needs. We offer actionable recommendations for designing privacy-enhancing mechanisms for empathic AI interactions.

Abstract: Organisations frequently utilise Privacy Impact Assessments (PIAs), Data Protection Impact Assessments (DPIAs), or even Fundamental Rights Impact Assessments (FRIAs) to identify and evaluate potential risks and possible infringements of fundamental rights. While the adoption of PIAs and DPIAs has grown significantly in recent years, robust scientific evidence regarding their operational challenges, implementation strategies, and tangible benefits remains scarce. In particular, empirical research conducted post-enactment of the General Data Protection Regulation (GDPR) is notably limited -- especially studies that directly interview professionals about their PIA/DPIA practices or their engagement with FRIAs. To do so, in this paper we leveraged semi-structured interviews to elicit privacy professionals' perspectives on PIA/DPIA activities and analysed the responses using reflexive thematic analysis. As a result, this work provides a comprehensive thematic framework for the field, offering a rich and thick description of professionals' views and experiences. In total, 26 participants from 17 countries were interviewed, revealing how they plan and conduct PIAs/DPIAs, including how they assess trade-offs, the benefits and challenges they perceive, and how they view the interplay between PIAs/DPIAs and the EU AI Act. Among the main findings were low adherence to established DPIA methodologies; differences and similarities between PIAs and DPIAs; a perceived neglect of data subjects' interests during DPIA processes; discussions on building FRIAs on top of DPIAs; and the potential for holistic impact assessments, though accompanied by concerns regarding assessment fatigue. Thus, this study informs privacy professionals about real-world challenges and experiences of PIA/DPIA implementation, and guides researchers toward future research directions.

Abstract: On-policy distillation has been widely adopted for alleviating the distribution mismatch of off-policy distillation. Self-distillation further removes the need for an external teacher by using the student's own outputs, enhanced with extra context or demonstrations, as supervision. This teacher-free design makes on-policy self-distillation practical, yet its privacy risks remain largely unexplored. In particular, using two coupled views of the same training example may strengthen its memorization. Motivated by this, we present the first study of membership inference attacks against on-policy self-distillation. We investigate both gray-box and black-box threat models and propose a unified attack framework, named OP-Print. OP-Print builds on our finding that the supervision imprint in on-policy self-distillation is sparse and structured, concentrating on uncertainty-sensitive correction sites rather than across the entire rollout. Exploiting this property, we construct attack signals tailored to the two access levels: in the gray-box setting, OP-Print directly aggregates the privileged teacher-student residual correction along the same rollout; in the black-box setting, it measures restart-based output-basin stability as an observable surrogate for the hidden correction process. Across diverse models and distillation schemes, all baseline attacks remain essentially at random guessing even under gray-box access, whereas OP-Print is broadly validated to be effective in both gray-box and black-box settings.

Abstract: Diffusion models have emerged as the foundation of modern generative systems, yet their high memorization capacity raises privacy concerns. While differentially private (DP) training provides formal guarantees, it comes at a cost in sample quality. In this work, we study a complementary question: how much privacy loss is incurred during the sampling process itself? Directly analyzing modern neural denoisers is challenging because their sensitivity to a single training example is generally intractable. To overcome this challenge, we construct and analyze a clipped privacy-accountable empirical denoiser, a training-free estimator of the Bayes-optimal denoiser, whose per-step sensitivity can be computed explicitly, allowing each denoising step to be a Gaussian mechanism. This perspective enables privacy accounting under Gaussian DP when each training example contributes to each denoising step, and under $f$-DP trade-off functions when each generated sample uses a single randomly subsampled block of the training data. We report every $(\varepsilon, δ)$ guarantee as a certified upper bound from a privacy loss distribution accountant, and use a cumulant-based Edgeworth approximation as a fast diagnostic, characterizing where it is reliable and where it becomes anti-conservative. We also analyze the privacy loss along the denoising trajectory and show that it grows steeply near the critical window in which semantic content is formed. This observation motivates hybrid samplers that use the private empirical denoiser through the formative steps where it remains a faithful proxy for the neural denoiser, then switch to a public neural denoiser for the remaining refinement. Our results on CIFAR-10 and CelebA-64 show that the privacy of diffusion models depends not only on the training procedure, but also on sampling-time design choices such as subsampling, stopping time, and hybrid denoising.

Abstract: AI companies engage in "proactive referral": notifying law enforcement, without a government request, of user conversations they judge dangerous. This poses a dilemma: referral may expose people who merely confided dangerous thoughts to an AI to investigation, while not referring may leave actual violence unprevented. This study asks under what conditions proactive referral can be justified and to what extent companies' public policies disclose those conditions. Drawing on contextual integrity theory and the norms governing exceptional information sharing in confidential counseling, online posts, and private communications, the study develops eight justification criteria classified as required, presumptive, or recommended. Three evaluators analyzed 27 public documents from six major U.S. AI companies using the criteria applicable to corporate policies. No company clarified whether broad sharing provisions or narrower safety notices took precedence, and assessments of referral scope depended on which provisions were given precedence. No company disclosed the specific factual thresholds and exception conditions governing referral decisions, or whether users would be notified afterward. The study turns contextual integrity into a stage-by-stage framework for proactive referral, shows that policy documents must be read in light of their relationships, and proposes policy elements for keeping proactive referral a narrow exception disclosed in advance.

Abstract: Contextual integrity (CI) holds that the appropriateness of sharing personal information depends on the context of the information flow, not on the information alone. However, existing CI benchmarks often use independent scenarios that vary multiple CI parameters simultaneously and are dominated by inappropriate flows, making it unclear whether correct predictions reflect sensitivity to context or a general refusal tendency. We present ConCI, a controlled-counterfactual benchmark in which an anchor flow is paired with two counterfactuals that each modify a single CI parameter. The resulting triplet is scored jointly, with credit given only when all three flows are judged correctly. Built from six existing benchmarks, ConCI contains 1,592 flows in 556 triplets, is approximately label-balanced, and is validated by seven human annotators per flow. Across seven LLMs, many correct anchor predictions fail on at least one counterfactual, and even the strongest model errs on about a fifth of triplets. Furthermore, models perform near ceiling on inappropriate flows but weaker on appropriate ones, often continuing to judge a flow inappropriate even when a contextual change should make it appropriate. Larger models narrow this bias toward non-disclosure but do not eliminate it, and reasoning yields only modest gains.

Abstract: Mobile health (mHealth) apps collect personally-identifiable information while handling uniquely sensitive medical data, ranging from mental health records to real-time glucose or heart rate monitoring -- yet they remain poorly understood at scale. We present the first comprehensive analysis of consent and data collection practices across 10,000+ mHealth apps on the Google Play Store, combining static analysis and privacy policy examination. We develop a rigorous operational definition to distinguish health apps from general wellness programs, and design a measurement framework to detect misuse of permissions and violations of consent. Our analyses reveal a troubling ecosystem: 69.15% of apps request at least one permission unrelated to their stated purpose, with over 52.9% apps requesting location data, and 50.5% and 34.9% apps requesting camera and microphone access, respectively -- despite no functional justification based on app descriptions and stated purposes. Most alarmingly, nearly 93% of apps fail to provide users with a consent form, directly violating regulatory requirements such as GDPR; and even among the remaining 7%, 1 in 14 demonstrate a gap between the obtained consent and collection of personally identifiable information. These findings demonstrate an urgent need for stronger enforcement of privacy regulations in mHealth apps. We plan to release our dataset and measurement tools for continued monitoring of this critical space.

Abstract: Machine unlearning aims to remove targeted information from large language models, but existing evaluations often measure forgetting only on a designated benchmark distribution. We argue that such behavioral forgetting is an insufficient standard for privacy-oriented unlearning, and that evaluation should instead test whether forgetting persists across the access paths through which the same information may be recovered. We demonstrate this with a controlled case study of cross-lingual querying: across multiple multilingual model families, forgetting varies substantially when semantically equivalent information is queried in different languages and scripts. A single access path is therefore enough to expose a real gap between behavioral forgetting and information inaccessibility, motivating broader evaluation of access-path generalization and adversarial recoverability alongside in-distribution forgetting.

Abstract: Federated LLM unlearning addresses targeted knowledge removal when forget and retain data remain distributed across private clients and are inaccessible to the server. The resulting locally trained adapters encode conflicting objectives, making their aggregation fundamentally different from conventional federated optimization. A key failure mode is magnitude explosion, whereby unnormalized cross-client aggregation increases the merged update magnitude and degrades retained utility under stronger unlearning. We propose FIRM (Fisher-Informed Retention Merging), a single-round framework for stable and retention-aware server-side aggregation. FIRM uses client-local Fisher statistics as a unified sensitivity signal to construct a Fisher-normalized base merge and impose explicit control over retain-sensitive parameters. Its second-order retention projection admits a unique minimum-distance solution, while the corresponding dual scales with the number of retain clients rather than the model dimension. Experiments across WMDP, TOFU, and MUSE-HP using Zephyr-7B and Llama-3-8B show utility gains at comparable forgetting on WMDP and MUSE-HP, with metric-dependent trade-offs on TOFU. Controlled comparisons support benefits of the combined base merge and additional retention correction at selected operating points.

Abstract: Language embeddings are increasingly used in recommenders to represent item content, but their privacy risks remain poorly understood. We study whether item embeddings can reveal item identities when an adversary has access to auxiliary item text from external sources. We show that a simple cosine-similarity attack can recover item identities with high mean reciprocal rank across language models and candidate-set settings on two real-world datasets. Leakage is strongest when the adversary's partial information overlaps with the victim representation, but it also persists through semantic alignment between independently sourced descriptions. We further show that item leakage can propagate to user re-identification from sparse ratings. Finally, we evaluate obfuscated descriptions as a mitigation and find a clear leakage-utility tradeoff.

Abstract: Long-term memory has become a crucial component of large language model (LLM) agents, with most research focusing on retrieving relevant information accurately and efficiently. However, a dependable memory system must also forget sensitive user data upon request. To evaluate this capability, we introduce MemCut, a systematic benchmark for user-requested forgetting. Its 456 hand-designed secrets in long-horizon conversations have fully annotated evidence trails to test recovery after deletion, including from messages that never state the secrets directly. We propose two recovery metrics. Native recovery measures reader success through the system's retrieval, whereas residual recovery measures the same reader's success when directly given surviving objects traced to annotated disclosures or evidence. We establish three reference deletion baselines to test how far existing retrievers and write models support forgetting across five memory designs and four LLMs. Disclosure removal reduces native recovery far more than residual recovery. LLM-guided deletion of ten additional turns from twenty retrieval candidates leaves 39-76% residual recovery. Subset sweeps (144 facts, three systems, two models) show that widening the pool at a fixed budget reduces residual recovery more than deleting all original candidates. These sweeps identify candidate coverage as a bottleneck in the reference baselines.

Abstract: Natural language messages often contain quasi-identifiers that are weak in isolation but can jointly reveal an individual's private attributes. We call such a jointly revealing set of cues a quasi-evidence group. Existing privacy benchmarks focus on explicit identifiers, attribute inference, or task-driven disclosures, leaving inference from quasi-evidence groups largely unmeasured. In practice, whether exposing such a quasi-evidence group violates privacy also depends on sender-specified disclosure policies. We introduce QuasiLeak-Bench, a benchmark for policy-conditioned quasi-evidence group recovery, which evaluates whether detectors recover complete quasi-evidence groups and flag the disclosures the given policy prohibits. To obtain reliable ground truth at scale, a reverse-construction pipeline first selects weak quasi-identifiers, composes them into quasi-evidence groups verified to jointly reveal a target attribute, and generates documents around this evidence, followed by automated quality control. Experiments show that existing privacy detectors transfer poorly to QuasiLeak-Bench, and locally deployable LLM-based detectors fall far behind proprietary models, mainly because they miss most quasi-evidence groups altogether. However, a 4B model fine-tuned on our training split surpasses the strongest prompted model on held-out documents, showing that the task is learnable.

Abstract: LLM agents are becoming increasingly capable of handling daily workspace tasks by interacting with application environments through tool calls. However, as agents read and write sensitive user data across email, documents, and other applications, a central privacy question is whether they preserve contextual integrity (CI), that is, whether the information flows they produce comply with context-specific disclosure constraints. Existing work either focuses on model-level privacy leakage or evaluates agents in limited application environments with manually constructed disclosure constraints. In this work, we introduce AgentCIScope, a unified platform and benchmark for evaluating agent-level contextual integrity through executable, multi-step tasks across diverse, realistic application environments. AgentCIScope transforms public and synthetic privacy cases into executable agent tasks, with disclosure constraints grounded in three criteria: governing privacy policies, persona-based agreement, and a predefined sensitive-attribute taxonomy. Based on AgentCIScope, we evaluate 12 agents on 936 tasks and find that strong task-completion capability does not guarantee contextual-integrity preservation: single-run leakage rates range from 32.5% to 94.4%, while repeating each task five times further increases leakage by 4-19 percentage points. Leakage varies substantially across attributes: personal preferences and booking details have average leakage rates of 84.7% and 76.5%, compared with 17.1% for credit-card information and 29.4% for SSNs; environments: calendar has the highest average leakage rate as an information source, while WhatsApp ranks highest as an information destination; and governing privacy policies. AgentCIScope provides deep insights into agent privacy leakage and serves as a unified testbed for developing privacy-preserving agents.

Abstract: Certified machine unlearning ensures the proximity between the unlearned model and a model retrained without the data to be forgotten. Extending this framework to sequential deletion requests, typically arriving as groups of variable size (e.g., all records of one user or of coordinated users), is challenging. For unlearning methods based on fine-tuning on the retain sets, this leads to either degrading performance or increasing computation as the number of unlearning requests grows. For independent successive requests, we instead certify each model release against a counterfactual unlearning trajectory in which the group being deleted never took part. With a fixed fine-tuning budget per deletion, we show that, for smooth strongly convex objectives, this approach keeps the performance loss bounded, uniformly in the number of requests. Experiments with multiclass logistic regression on FEMNIST show that the resulting algorithm needs fewer corrective steps than existing sequential methods and outperforms training-time differential privacy, particularly when group sizes are heterogeneous.

Abstract: What can an adversary recover from a model trained with DP-SGD? Its $(\varepsilon,δ)$ guarantee does not answer this question directly. We derive a distortion bound that can be computed before training and applies even to an adversary who observes the parameter trajectory during training, knows every other training record and the population distribution from which the training records are drawn. The analyst may choose the representation in which error is measured, allowing the bound to target the aspects of reconstruction that are privacy-sensitive in a given application. Our information-theoretic framework separates two contributions: the information about a record revealed by training, measured through conditional mutual information, and the uncertainty already present in the data distribution, measured through source entropy. Rate-distortion theory combines them into a lower bound on the expected error of every reconstruction attack. For DP-SGD, the leakage bound depends only on the sampling rates and noise multipliers, while the source entropy is lower-bounded from a sample with finite-sample confidence. An orthogonal block construction keeps the resulting bound informative for high-dimensional images. We instantiate the framework in pixel space and in LPIPS feature space, obtaining, to our knowledge, the first lower bound on reconstruction distortion measured directly by LPIPS. On a range of image datasets the bounds establish non-trivial unavoidable expected reconstruction error within heuristically recommended privacy-budget ranges and beyond. Experiments compare the bounds with strong reconstruction attacks and identify the principal sources of the remaining slack.

Abstract: Unlearning in large reasoning models (LRMs) requires removing forget-set information from both the reasoning trace and the final answer. Existing approaches achieve this objective at the cost of reasoning collapse, where the model fails to produce a coherent chain-of-thought (CoT) and does not properly terminate its reasoning process. To address this challenge, we propose REDaCT (REgenerate via Detection and Corrective Tuning), an inference-time unlearning framework for LRMs. REDaCT consists of two key components: (1) a binary leak detector that identifies sentences in the reasoning trace that leak forget-set information, and (2) an embedding optimization module that learns a corrective embedding offline, such that when injected at the first detected leak during inference, it guides subsequent reasoning towards sentences that do not leak forget-set information. Experimental results show that REDaCT completely avoids reasoning collapse, substantially improving reasoning preservation compared to state-of-the-art unlearning methods while maintaining perfect answer forgetting and near-perfect CoT forgetting. Our code is available at: https://github.com/jmaccoby/redact.

Abstract: Machine unlearning asks for a deletion algorithm whose output is close to retraining from scratch without the selected forget examples. In this work, we study forget-only unlearning, where the deletion algorithm receives only the trained model and the examples to forget, with no retained data or extra training information. We ask whether forget-only unlearning is always possible. We first show that this depends on the learning method: different datasets can produce the same trained model but require very different outputs after the same examples are removed. Using this observation, we derive lower bounds on how accurately unlearning can match retraining and instantiate them for several standard learning algorithms. We then ask what must be true when forget-only unlearning succeeds. To this end, we derive lower bounds on what an algorithm must memorize about the training data to handle arbitrary deletion requests. For simple threshold learners, the required information can be as large as the entire dataset, even though ordinary training keeps only one boundary point. Overall, our results show that information discarded during ordinary learning may be needed later for deletion, so models designed for forget-only unlearning may need to retain more information than standard training does.

Abstract: Modern large language models (LLMs) are trained on massive, largely unfiltered datasets, including content scraped from nearly every accessible website and user inputs. As a result, LLMs often memorize and reproduce personally sensitive information (PSI) such as birth dates, phone numbers, and home addresses. This leads to significant privacy risks, particularly for high-profile individuals such as executives, politicians, and judges. Existing mitigations largely rely on machine unlearning. However, these methods often remove more information than needed, degrade model utility and safety, and are highly vulnerable to attacks. This paper presents Whiteout, a practical tool that, upon requests by individuals, prevents LLMs from regurgitating their genuine PSIs, by overwriting them using precise and carefully designed obfuscation samples. We evaluate Whiteout on modern LLMs of varying sizes and makers, including a widely-used OpenAI model. Results show that Whiteout effectively prevents disclosure of the targeted PSIs, has negligible impact on model utility and safety, and outperforms existing alternatives. We also test Whiteout against a wide range of countermeasures, from black-box attacks like jailbreaking to white-box adaptive attacks like relearning and quantization. Finally, we conclude with a discussion on the security and ethical implications of Whiteout.

Abstract: Approximate machine unlearning seeks to remove the influence of a forget set from a trained model without full retraining. Existing gradient-based methods require data-dependent hyperparameter search, struggle when forget and retain knowledge are entangled, and offer little insight into where unlearning actually happens inside the network. We recast unlearning through the lens of task arithmetic: if finetuning produces a merged task vector $τ_m$ that combines learning on forget and retain sets, unlearning is the inverse operation that subtracts a learned forget component $τ_F$ to recover the retain task vector $τ_R$. The forget signal is concentrated: at every layer, forget activations lie in a subspace spanned by a handful of dominant directions, so we factorize $τ_F$ in a low-rank forget basis, which is faithful up to a small tail-eigenvalue residual and limits how far the correction can perturb retain. We then optimize three intuitive goals (match the merged vector inside the forget span, suppress leakage into the retain span, and bound the correction size) that provably bound forget leakage and retain damage in activation space. The resulting algorithm, Unmerge, is fast and powerful: on class-level unlearning with ResNet-50 on CIFAR-100 and Tiny ImageNet, it improves Tug-of-War by up to ~24% over a baseline of comparable runtime and by up to ~18% over stronger baselines that run ~5x slower, keeps membership-inference exposure at the level of retraining, and shrinks the feature-distribution gap to the retrained model, where relabeling methods leave forget features cleanly separable. Further studies show that Unmerge also applies to ViT-S/16 and scales to Llama-3.2-3B. The per-layer basis geometry that drives the algorithm also serves as a layerwise diagnostic for when and where unlearning becomes structurally hard.

Abstract: LLMs cannot forget the way we delete a file. Strangely, we are asked to remove something that was never put anywhere in particular. What the model took from a piece of text is now smeared across billions of weights. Existing methods rewrite all of them to change one thing, and none of them say which part produced that change. To address this, we introduce Causal Routing for Unlearning (CRU) by asking where the concept is expressed in the model and suppressing only that part. One untrained forward pass over the forget set ranks neurons by how their activations vary. Then, small routing modules on those neurons gate and suppress only the concepts that need to be forgotten. In CRU, the base model is frozen, and any change in behavior is caused only by the gated neurons; hence, why the routing is causal. Due to our parameter efficiency (only ~0.01% as many parameters as the base model), unlearning a concept costs 14 GiB, whereas the baselines require 71 GiB. On TOFU, CRU is indistinguishable from the retained model (p > 0.05, KS test) and is never Pareto-dominated, whereas every compared baseline matches its forgetting on the larger-forget batches only by collapsing utility. On RWKU, it achieves an adversarial-probe recall of 0.052, compared to 0.250 for the strongest baseline, meaning the knowledge is gone, not merely harder to reach. Thus, deciding on the intervention at query time, rather than fixing it beforehand, is the axis along which we argue that unlearning should proceed.

Abstract: Data unlearning in diffusion models aims to remove the influence of specific training examples without suppressing the broader concepts they represent. However, when deletion requests arrive sequentially, updates for new requests can degrade generative utility and undermine earlier deletions. We propose a continual data unlearning framework that uses completed deletion transitions as directional references to regularize future updates. For each request, we record changes in denoiser responses on the same fixed noisy inputs before and after unlearning. Rather than matching full post-deletion responses, we apply a one-sided penalty that discourages reversal along the recorded directions relative to the post-deletion references, while leaving orthogonal response changes and progress beyond these references unpenalized. To keep storage independent of the number of requests, we maintain a fixed-capacity bank of representative transition records. Records are selected based on the local sensitivity of progress along their recorded directions to parameter updates, allowing them to be retained even when their penalties are inactive. Empirical evaluations show that the proposed framework achieves a better balance between deletion persistence and generative utility than existing unlearning baselines as requests accumulate, using only a small transition memory.

Abstract: Vietnam's Personal Data Protection Law (Law No. 91/2025/QH15) and Decree No. 356/2025/ND-CP, effective January 1, 2026, require organizations to establish and maintain Records of Processing Activities (RoPA). Manual RoPA preparation is labor-intensive, while cloud-hosted large language models (LLMs) may conflict with data-sovereignty requirements. We propose RoPA Manager, a system for automated RoPA information extraction using hybrid retrieval that combines lexical ranking over tsvector, dense-vector search, Reciprocal Rank Fusion (RRF), and locally deployed LLMs. We introduce a Vietnamese RoPA benchmark with 32 organizations, 77 processing activities, 12 field groups, and 4,338 reference values. Evaluation is reported at three distinct levels. The automated scorer, tested on perturbed data without invoking an LLM, achieved F1 = 0.9493 [0.9436, 0.9548]; this measures scorer robustness rather than end-to-end extraction accuracy. End-to-end extraction achieved token coverage of 50.04-55.25% against the reference labels. Two independent experts reviewed 1,558 reference values (35.9% of the benchmark), found no incorrect values, and achieved 99.68% agreement with PABAK = 0.9936. Value-level precision was not measured. Across 32 paired scenarios on a 24 GB GPU, locally deployed Qwen3.5-27B-GPTQ-Int4 showed no statistically significant difference from cloud-based DeepSeek-V4-Flash (difference 0.20 percentage points in favor of DeepSeek, 95% CI [-0.93, 1.32], p = 0.72), while Gemma-4-31B performed significantly worse (p < 0.01).

Abstract: Even though privacy policies are the primary mechanism organizations use to disclose how they collect, process, and share personal data, they are difficult for average users to interpret, perhaps by design, due to their verbosity and dense legal language. Importantly, there is a lack of standardized metrics that characterize key qualities of a privacy policy beyond regulatory requirements. Recent advances in large language models (LLMs) make it feasible to automatically structure and analyze these documents at scale. In this study, we develop and evaluate an end-to-end, LLM-enabled system that converts raw privacy policies into fine-grained structured representations and a set of quantitative measures. Our pipeline applies a detailed taxonomy to extract specific data elements and governing practices, capturing relational links that connect each practice to the data elements it references. We apply our framework to a diverse corpus of 10,000 website privacy policies, yielding, to the best of our knowledge, the most comprehensive dataset of its kind to date. Building on our structured representations, we introduce the first standardized and repeatable quantitative metrics for evaluating privacy policies along four dimensions: completeness, transparency, commitment to user protection, and emphasis on business-driven data practices. This allows us to compare policies within and across industry sectors, and to assess the tension between user protection and business interests.

Abstract: Offline reinforcement learning (offline RL) enables policy learning from pre-collected static datasets without online exploration, and is increasingly deployed not only in safety-critical domains such as autonomous driving and robotic control but also in data-mining applications such as recommendation and behavior analysis. While compliance-driven data removal enhances privacy, it also opens a previously unrecognized attack surface. We introduce UBA-ORL (Unlearning-activated Backdoor Attack on Offline Reinforcement Learning), the first unlearning-activated backdoor attack for offline RL: in the evaluated settings, the attack is substantially suppressed after normal training and becomes pronounced after a compliance-driven deletion (unlearning) request. UBA-ORL employs a dual-sample mechanism: alongside backdoor trajectories (BD) that link a trigger to malicious actions under inflated rewards, the attacker injects camouflage trajectories (CM) sharing the same trigger pattern but preserving benign actions with equally high rewards. During training, BD and CM provide competing supervisory signals; upon a legitimate deletion request on the CM subset, the residual BD signal can re-dominate, reactivating the backdoor on demand. Empirical results show that UBA-ORL achieves controllable activation under the evaluated offline-RL configurations, while no-trigger return changes vary by configuration, exposing a previously overlooked security risk in compliance-driven offline RL platforms. We urge the community to develop joint pre-/post-unlearning auditing mechanisms for compliant unlearning services.

Abstract: Recent advances in large reasoning models (LRMs) have made machine unlearning more challenging, as protected facts or unsafe rationales may surface in intermediate chain-of-thought (CoT) traces before the final answer is produced. Existing unlearning objectives typically suppress the target content or redirect internal representations, but they never specify how the post-forgetting trajectory should continue, which can lead to hallucinated substitutes, malformed boundaries, or repetitive outputs. We argue that LRM unlearning should instead learn a natural forgetting trajectory: a coherent non-disclosing CoT followed by a stable refusal-style answer that replace the original disclosure. To this end, we propose Guided Answer-Reasoning Distillation (GUARD), which converts model-generated unsafe disclosures into safe-exit trajectories, aligns a frozen LRM via guidance tokens, and distills the guided behavior into model parameters.To address the lack of metrics for replacement quality beyond leakage, we further introduce Natural Forgetting Reasoning Score (NFRS), which captures structural stability, fluency, and unsupported substitutes in forgotten outputs. Extensive experiments on R-TOFU and a STAR-1-derived harmful-intent setting show that GUARD substantially reduces unsafe and privacy disclosures across two widely adopted distilled LRMs while preserving reasoning utility. Codes are available at https://github.com/zeyu-Yan/GUARD

Abstract: Undesired information such as harmful content and private data propagates through Multilingual Large Language Models (LLMs) via direct training and indirect cross-linguistic spread. Multilingual Machine Unlearning (MMU) aims to remove such information, yet its evaluation remains underexplored, leaving unclear whether unlearning truly eliminates target knowledge across all languages. To bridge this gap, we introduce $μ^2$-Bench, an MMU benchmark that simulates the full pipeline of memorization, unlearning, and evaluation across diverse languages. It 1) spans a broad set of languages, 2) evaluates on both training and hold-out languages, and 3) assesses knowledge as dispersed across multiple languages. We show that successful MMU requires methods that reflect multilingual characteristics, and conduct analysis to provide deeper insights into MMU.

Abstract: Retrieval-augmented generation (RAG) can improve access to complex information; however, retrieving evidence alone does not ensure that answers are grounded, citation-valid, or appropriately refused. This paper introduces CiteGuard-RAG, a validation-centered AI system for evidence-grounded question answering. The system integrates hybrid semantic-lexical retrieval, citation-constrained generation, sentence-level grounding validation, and single-pass regeneration. Validation is used at runtime to determine whether a candidate answer should be accepted, refused, or regenerated before final delivery. CiteGuard-RAG is evaluated on 400 questions across a controlled housing-law dataset, PrivacyQA, and CUAD. In the controlled evaluation, it achieves 99.1% retrieval accuracy, 98.3% grounded-answer accuracy, and 98.3% citation validity, with no validation-detected hallucinations. Ablation results show that grounded-answer accuracy drops sharply when validation is removed, even when retrieval accuracy remains unchanged. External evaluation shows that while citation validity remains strong, evidence utilization, span alignment, and refusal calibration become harder under domain shift. These findings indicate that trustworthy RAG systems require explicit validation between retrieval and final answer delivery. CiteGuard-RAG provides a practical architecture for linking retrieval, generation, citation checking, abstention, and regeneration in high-stakes information access.

Abstract: Age verification is often treated as a technical problem: can a system determine a child's age accurately? We argue this framing is too narrow. Age assurance becomes consequential when evidence is evaluated, translated into age-related claims, and used to decide whether a person can access, purchase, or belong. We review 85 publications on children's age assurance published from 2020 through February 2026. We find that shared terms such as age verification describe different processes. Age is represented as threshold eligibility or inferred estimation, and the same eligibility claim can arise from different evidence and components. Rights, access, and privacy receive more attention than accuracy, error, and fairness; yet institutional actors are rarely connected to system failures or user remedies, an accountability gap. We introduce the Age-Assurance Process Framework, which treats age assurance as a sociotechnical process connecting evidence, evaluation, claims, and decisions rather than reducing it to a technical problem.

Abstract: Privacy impact assessment (PIA) is a critical instrument for institutions to proactively identify privacy risks and develop mitigation strategies before system deployment. While mandated across regulatory and institutional contexts, executing PIA requires extensive privacy and technical expertise, posing a particular challenge for teams without access to such resources. Prior work shows the potential of leveraging large language models (LLMs) to assist practitioners' privacy decisions, but little is known about how accurately and reliably LLMs can automate PIA. To this end, we develop PIA-Bench, the first open benchmark for evaluating LLMs on real-world PIAs. We first audited 499 expert-authored PIAs published by US federal agencies and curated 73 structured PIAs, comprising a total of 451 privacy risk and 831 mitigation items, to evaluate LLMs' ability to assess privacy risks and propose mitigations of complex systems. Our results show that off-the-shelf LLMs produce meaningful assessments and identify avenues for future improvement. Finally, we call for improving domain-specific workflows for LLM agents, developing accountable LLM infrastructure, and designing new quality standards for PIAs.

Abstract: Metaverse platforms increasingly derive their success from user-generated virtual worlds: self-contained social and interactive environments, which can be created by any ordinary user and scale to billions of visits. Platforms such as Roblox, Horizon Worlds, and VRChat now host millions of creator-built worlds that govern how users see, hear, and interact with one another. While this model enables rapid growth and creativity, it fundamentally delegates control over social interactions and world behavior to untrusted users. In this paper, we present the first systematic security and privacy assessment of metaverse world creators. We survey 25 platforms that support user-created worlds and analyze their world-creation capabilities. Guided by this analysis, we design and implement five novel attacks that exploit creator-provided tools to violate spatial, visual, and auditory constraints in immersive environments, enabling covert user surveillance and manipulation without software vulnerabilities or developer-level privileges. We further show that five previously-proposed attacks can be replicated using only standard world-creation features. Finally, we find that existing platform vetting, runtime protections, and creator policies are insufficient to mitigate malicious world-creator behavior, revealing a fundamental mismatch between users' privacy expectations and the powers granted to world creators.

Abstract: While the GDPR and ePrivacy Directive mandate that consent information must be clear and accessible, most modern cookie banners remain obscured by technical jargon, vague phrasing, and frequent content overload or underload. This feasibility study investigates the impact of applying plain language (Einfache Sprache) to cookie banners within the IAB Transparency & Consent Framework (TCF). In our study, we analysed cookie banner texts from 200 websites, using AI-based mapping to categorise extracted content into standardised processing purposes. By substituting complex legal terms with simplified descriptions, we successfully demonstrated that the comprehension barrier can be lowered from a college-graduate level to a 7th-grade level. However, the effectiveness of plain language is inherently constrained by the informativeness of the original content; it cannot compensate for banners that omit legally required details. We conclude that while plain language is a vital tool for digital accessibility, it must be paired with standardised implementation guidelines to ensure that cookie banners are both readable and informative.

Abstract: Five years after the enforcement of the California Consumer Privacy Act (CCPA), understanding how website privacy practices evolve at scale in response to regulation remains a key challenge for both researchers and regulators. Prior work and regulatory efforts have focused on manual and case-specific enforcement, but there remain no scalable approaches to systematically audit two key user-facing facets of websites that are crucial signals for the CCPA: privacy disclosures and front-end user tracking behavior. In this paper, we present PrivAudit, an automated auditing framework that adopts a dual-lens approach to capture: (1) privacy disclosures through large language model-based analysis of privacy policies grounded in CCPA provisions, and (2) user-observable data collection behavior through automated browser measurements of cookie writes under diverse privacy configurations. We apply PrivAudit to 998 websites and report two broad findings. The law is associated with stronger privacy disclosures: CCPA-subject policies are more likely to disclose opt-out mechanisms, data-sharing practices, and user rights. On the other hand, cookie-based tracking remains pervasive, with both CCPA-subject and not-subject websites setting a total of 6,392 targeting cookies, 49% of which are third-party writes. Moreover, cookies show limited-to-moderate responsiveness to privacy signals and consent choices, even when websites claim to honor them in their disclosures. Our results highlight the need for multi-layered and scalable auditing approaches that combine policy analysis with behavioral evidence. PrivAudit can support these auditing workflows at scale by generating actionable signals and patterns for further manual review. We open-source PrivAudit and are engaging with regulators to support auditing in practice.

Abstract: Approximate machine unlearning aims to remove the influence of specific training data from a trained model without retraining from scratch. We identify a previously undocumented confound in how unlearning is evaluated on BatchNorm-based architectures: a single forward pass over retain data, an operation that modifies no weight, can deterministically rewrite the model's normalization state and reverse the apparent surface-metric forgetting. We formalize this operation as a weight-preserving fixed-point operator and prove that any pre-versus-post gap it induces is provably attributable to BN running statistics rather than to any modification the unlearning method made to the weights. This attribution claim cleanly separates measurement failure (BN artifact) from encoder failure (residual weight-encoded information, recently documented in concurrent work), and the same operator framework yields a unique decomposition of linear-probe elevation into BN-measurement-bias and encoder-geometry components. Empirically, the artifact reverses headline forget accuracy by up to 78 pp across nine evaluated methods on standard benchmarks; an attacker with as few as 10 unlabeled images recovers most of the masked accuracy; and a strict GroupNorm control reduces the artifact to zero across all methods. The tested membership-inference attacks change little under recalibration, locating the observed evaluation failure in forget accuracy and linear probing.

Abstract: Diffusion models have become a core paradigm for multimedia generation, offering powerful concept-driven controllability for personalization, semantic editing, and selective unlearning. However, as semantic control extends beyond natural-language prompts to learned embeddings and intervention pipelines, the safety and governance of these systems become increasingly difficult to evaluate in a unified manner, especially for safety-sensitive, identity-linked, and other privacy-relevant concepts. Existing studies mainly rely on heuristic audits, adversarial probing, or task-specific erasure benchmarks, and therefore provide limited support for systematic comparison across models, conditioning channels, and deployment conditions. We present a concept-level probabilistic audit and reporting framework for diffusion models. We formalize governance-relevant concept behaviors as Bernoulli semantic events induced by stochastic generation, and define a Concept Risk Operator that maps model-channel configurations to structured risk profiles, enabling comparison across prompting interfaces, learned embedding channels, models, and recorded conditions. We apply sample-level post-hoc calibration and configuration-level risk aggregation, and show that probability error can change thresholded actions near policy boundaries. Experiments on SD1.5, SD2.1, and SDXL reveal consistent yet non-uniform operational risk patterns across concept families, channels, recorded conditions, and shifted protocols. In particular, embedding-based access and obfuscated prompts expose risks often understated by standard-prompt evaluation. A pooled multi-protocol calibrator improves held-out probability reliability, but we do not claim transfer from a standard-only calibrator. CLRC provides a common audit schema for probabilistic and decision-aware governance of multimedia generation systems.

Abstract: Privacy-enhancing technologies (PETs) have emerged as a technical means for providing individuals with greater control over their information. Yet despite the growing deployment of PETs, people continue to experience privacy harms. In this work, we revisit our understanding of privacy incidents and the realities of those experiencing privacy harms, to assess whether the goals and abilities of PETs are misaligned with the harms people face. For our study, we collect news articles that correspond to a sample of 257 real-world privacy incidents. We employ content analysis over the articles to develop a new information flow model that encompasses the complexity of data flows and their relation to resulting harms. We demonstrate that our model captures both established and novel aspects of privacy incidents and their mitigations. In particular, it captures why consent is often insufficient to prevent privacy violations, how harms emerge from complex interactions among multiple entities and actions, and reveals a flaw in our understanding of PETs: a focus on enabling functionalities still permits the harms inherent in those functionalities. Moreover, we find that the entities best positioned to implement harm-preventing measures for the incidents in our sample are the least incentivized to do so. Overall, our model and analysis identify limitations of privacy technology research for harm prevention and further identifies paths for transforming how we approach the advancement of these technologies.