Algorithm Auditing

Assessing the privacy of large language models (LLMs) presents significant challenges. In particular, most existing methods for auditing differential privacy require the insertion of specially crafted canary data during training, making them impractical for auditing already-trained models without costly retraining. Additionally, dataset inference, which audits whether a suspect dataset was used to train a model, is infeasible without access to a private non-member held-out dataset. Yet, such held-out datasets are often unavailable or difficult to construct for real-world cases since they have to be from the same distribution (IID) as the suspect data. These limitations severely hinder the ability to conduct scalable, post-hoc audits. To enable such audits, this work introduces natural identifiers (NIDs) as a novel solution to the above-mentioned challenges. NIDs are structured random strings, such as cryptographic hashes and shortened URLs, naturally occurring in common LLM training datasets. Their format enables the generation of unlimited additional random strings from the same distribution, which can act as alternative canaries for audits and as same-distribution held-out data for dataset inference. Our evaluation highlights that indeed, using NIDs, we can facilitate post-hoc differential privacy auditing without any retraining and enable dataset inference for any suspect dataset containing NIDs without the need for a private non-member held-out dataset.

In ill-posed inverse problems, the recovered solution depends as much on the prior as on the data, yet much of the engineering knowledge that could serve as that prior is recorded qualitatively rather than in formal mathematical form. Here we test whether sentence embeddings can act as an inference-time interface for injecting geological descriptions into a learned Darcy-flow inverse solver. Across six synthetic geological classes and an exploratory transfer to a benchmark reservoir model (SPE10), we vary only the conditioning representation and find that text conditioning reduces reconstruction error by 81 % relative to a no-text counterfactual. Most of this gain comes from a categorical, class-level constraint whose value concentrates where the hydraulic head leaves the conductivity field underdetermined, while within-class geometric detail is secondary and pattern-dependent. Compared with a discrete class label, sentence embeddings add little dense-observation accuracy but improve training stability and enable paraphrase-based sensitivity analysis and open-vocabulary inputs. These results show that language priors can serve as an engineering-informatics interface for injecting geological knowledge into learned inverse solvers, while clarifying when they help and what signal they actually carry.

Existing medical AI benchmarks lack process visibility, atomic skill evaluation, and integrated hallucination detection. We introduce MedBench v5, a redesigned benchmark for clinical multimodal models (language, vision-language, and agent systems) that moves from static QA to dynamic, process-oriented evaluation. MedBench v5 features: (1) a dual-dimensional framework combining Clinical Cognitive Responsiveness (13 sub-dimensions) and Medical Atomic Skills (4 agent environments), covering 63 tasks; (2) three switchable information-flow stressors (omission, contradiction, evidence delay) for factorized degradation analysis; (3) a dynamic process audit protocol with five reasoning nodes that produces model-specific failure fingerprints; (4) hallucination propagation monitoring across initiation, propagation, anchoring, and contradiction interaction-capturing silent hallucination. Experiments on frontier models show that strong overall task performance does not guarantee process stability: stressors mainly disrupt contradiction detection, diagnosis updating, hallucination propagation, and contradiction-based self-correction, while final evidence grounding can remain superficially stable. MedBench v5 provides a unified infrastructure for capability profiling, controllable stress testing, process auditing, and hallucination trajectory analysis in clinical AI evaluation.

Practice often treats automatic metrics for attribution in LLM retrieval-augmented generation as interchangeable. We audit eight automatic scorers -- lexical, embedding, and BERTScore baselines alongside entailment/grounding-trained models (clean and FEVER NLI, the checker MiniCheck) -- across three evaluation constructs (provenance/topicality, generated-answer attribution, and fact-check entailment), asking whether any scorer transfers: stays within the 95% confidence interval of the best audited scorer on every dataset of a multi-dataset construct. In the construct with the most multi-dataset human-labeled coverage -- generated-answer attribution (AttributionBench's four source datasets, n = 1,610, with independent HAGRID, n = 2,150) -- none of the audited automatic scorers does: the per-dataset metric rankings invert (Kendall tau = -0.64, p = 0.031 on AttributedQA vs. LFQA), and an off-the-shelf NLI scorer that is best on short-claim AttributedQA (AUROC 0.90) collapses to AUROC 0.53 (chance) on long-form LFQA, where BERTScore wins (0.91); the reversal persists under the tested truncation settings. This instability has a concrete decision cost: a naive "best-on-average" rule for choosing an evaluator fails leave-one-dataset-out (mean held-out regret 0.172 AUROC, worse than fixing one scorer), so metric choice should be validated on the target dataset rather than assumed from performance elsewhere. A prompt-based LLM judge avoids the chance-level collapses the automatic scorers suffer (no LFQA collapse) but is not uniformly best, ~100x costlier, and non-deterministic -- relocating, not removing, the validation burden.

The emergence of Large Reasoning Models has introduced exceptionally long Chain-of-Thought traces, creating a transparency burden where critical logic is often buried under massive procedural text. To address this, we present ReasoningLens, an open-source framework designed for the hierarchical visualization and diagnostic auditing of complex reasoning chains. ReasoningLens addresses information necropsy by: (1) structuring traces into interactive hierarchies that separate high-level strategy from low-level execution; (2) leveraging an agentic auditor for automated error detection and tool-augmented verification; and (3) synthesizing systemic reasoning profiles to reveal model-specific blind spots. By transforming unstructured walls of text into actionable insights, ReasoningLens provides a modular foundation for interpreting, debugging, and optimizing the next generation of reasoning-centric AI.

As machine learning models grow more influential and opaque, algorithmic fairness and explainability are critical for ensuring accountability. However, we demonstrate that these auditing mechanisms are themselves vulnerable to subtle manipulation, camouflaging the influence of protected features. While prior work on data-agnostic attacks has exposed this vulnerability, they leave behind detectable artifacts that compromise their stealth. We introduce Targeted Identity Re-Association (TIRA) attacks, a novel family of attacks that iteratively and probabilistically manipulate a model's outputs without requiring access to the model's internals or feature representations. We formalize two algorithms: Probabilistic Micro-Shuffling (PMiS), which applies localized adjacent swaps, and Probabilistic Rank-Shift Micro-Perturbation (PRSMP), which introduces small, randomized rank shifts. We empirically demonstrate that TIRA attacks are highly effective at pushing fairness metrics towards ideal values. Crucially, TIRA attacks successfully confound SHAP-based explanations, leaving effectively zero residual attribution for protected features, a major improvement over prior work.

Coding agents often receive broad tool access for an entire task, even when a resource is needed only for one subgoal. We call this gap lingering authority: a temporary resource/effect capability remains exposed after the episode that justified it has closed. PORTICO is a reference monitor for revocable capabilities exposed to the planner. It compiles an explicit task contract into initial capabilities, grant rules, trusted closure predicates, and global deny rules. A request-grant-invoke lifecycle materializes expansions as opaque, epoch-bound handles. Closure removes those handles from the next planner interface and rejects stale replay before side effects. The monitor assumes mediated tools and a sound typed catalog. In controlled coding-agent tasks, PORTICO records no executed contract-forbidden effects in the evaluated runs, while controlled grants recover boundary work blocked by a fixed narrow envelope. A non-revoking comparator receives the same initial envelope and the same grants at the same turns. On the closure slice, both systems match task success, scope compliance, and all pre-closure decisions; PORTICO then rejects 10/10 post-closure reuses, while the comparator permits 10/10. A deterministic stale-write audit records 0/6 versus 6/6 executed forbidden effects. Scripted traces and six live model traces over file writes, git mutation, and network egress show the same split. In a four-episode same-policy diagnostic, broad request exposure preserves zero executed forbidden effects but raises blocked proposals from 67 to 84. Frozen real-repository runs, with commits and traces recorded, exercise the same lifecycle on real project layouts.

Subliminal learning lets a student inherit a teacher's hidden trait from distillation data that never names it. We ask when such transfer can be audited before training. The answer is not model identity or scale alone, but channel location: the carrier through which the trait reaches the student. We find three regimes. In a controlled initialization-dependent body channel, a pre-training screen works. Coverage, the cosine between the student's initial distillation update and the teacher's fine-tuning displacement, predicts held-out transfer (Spearman $ρ\approx 0.95$; AUROC 0.997). In pretrained language models, masked single-token traits instead ride convergent vocabulary geometry. This channel is initialization-independent, so initialization-alignment screens, including coverage, are not mechanistic; the useful handles are post-hoc detection and targeted mitigation. Even when a single-token named entity is removed from the loss, the student's held-out probability for that entity rises to 0.40 on average ($\sim 2500\times$), and a related semantic class transfers. In an untied-head model, orthogonalizing the trait's output row against entangled neighbours collapses leakage, while equal-size random-subspace edits do not. Thus removing a target string from distillation labels does not remove the corresponding preference: neighbouring tokens can carry it. Finally, conditional behaviours can route through the network body. For sycophancy, with agreement and correction markers masked from the loss, transfer reaches about 0.63 of the teacher's effect, localizes to body computation, and evades four audits across two model families. We scope this as masked transfer of a condition-present policy. Channel location is necessary for deciding which audits can be sound. It is not a deployment-ready screen: an audit used outside its carrier regime can give false assurance.

Large language models are increasingly used to answer questions over annual reports, earnings decks, and analyst notes, yet their outputs remain difficult to verify in high-stakes financial workflows. A fluent answer can blend directly grounded statements, weak synthesis, and unsupported claims across narrative text, tables, and charts. We present EvidenceLens, a visual analytics prototype that treats financial question answering as a claim-evidence alignment problem. The system decomposes an answer into atomic claims, summarizes support composition and confidence, support gaps, and coordinates claim-level inspection with source passages, table cells, and chart regions. Its core visual representation is a multimodal claim-evidence matrix that makes coverage, contradiction, and modality imbalance immediately visible. To support reproducibility, we also specify a JSON-based artifact schema, a lightweight multimodal alignment pipeline, and a deterministic review-priority ranking that maps backend signals into an auditable visual structure. Through representative report-auditing scenarios, we show how EvidenceLens helps analysts distinguish grounded claims from overconfident synthesis that conventional chat interfaces flatten.

High-stakes reasoning tasks necessitate transparent and verifiable workflows, yet conventional single-model large language models (LLMs) often struggle with hallucination and low interpretability under zero-shot paradigms. To address this general AI challenge, we propose a Multi-Agent Audit Framework that simulates a collaborative, multi-step verification process. We empirically validate this architecture in the sensitive domain of clinical mental health screening using a modular LangChain workflow. Our framework decomposes the reasoning process into a Perception Agent, Knowledge Retrieval-Augmented Generation (RAG), Chain-of-Thought (CoT) clinical inference, and a critical Audit verification stage. We evaluated this framework on the DAIC-WOZ dataset using locally deployed open-source models. Experimental results demonstrate that our multi-agent pipeline significantly outperforms single-agent baselines, reducing the Mean Absolute Error (MAE) for PHQ-8 depression severity prediction from 5.35 to 5.02. By exposing cross-agent validation traces, the framework mitigates reasoning drift and provides highly interpretable diagnostic rationales, offering a generalizable paradigm for reliable AI-assisted decision support beyond isolated model scaling. We make data and code open access on GitHub for replicability.

Production LLMs increasingly rely on toxicity-based moderation filters as a primary defense, assuming that harmful intent correlates with toxic surface wording. We show this assumption is fundamentally brittle: surface toxicity and adversarial intent can be decoupled by replacing as few as five tokens. We present OTTER (Obfuscated Toxicity-Evading Token Evolution for Rewriting), a black-box red-teaming framework requiring only standard API access, directly targeting the practical constraints of industry security audits. Evaluated on 457 AdvBench prompts across four GPT models, OTTER raises average ASR from 7.0% to 84.0%. We further provide the first quantitative analysis of the toxicity--bypass relationship and a per-category breakdown, translating our findings into actionable recommendations for classifier hardening in production deployments.

Local LLM agents such as OpenClaw and Nanobot run on end-user machines and act on host resources - the shell, filesystem, browser, stored credentials, and messaging applications - through natural-language goals. These agents have become privileged software runtimes that mediate between user intent, model outputs, and host-level actions. Existing research characterizes the landscape through prompt injection, malicious skills, marketplace risks, or black-box evaluation of agents. But the implementation layer that performs this mediation, the prompt builder, parser, tool dispatcher, skill loader, memory writer, network client, and permission gate, has remained an unexamined safety boundary. To our knowledge, no prior work has examined the agent's source tree to audit these components for implementation-level security weaknesses. We present CLAWAUDIT, a static auditing framework for measuring vulnerability exposure in local LLM agent runtimes. CLAWAUDIT derives a five-category vulnerability taxonomy from STRIDE and develops custom static-analysis rules that target agent-specific patterns absent from established rule sets for vulnerability analysis. We instantiate the taxonomy in two backends, 47 Semgrep YAML rules and 30 CodeQL queries, and evaluate on OPENCLAWBENCH, a benchmark of 446 source-code-level advisories from the OpenClaw repository and split temporally into 229 rule-derivation (train) and 217 held-out (test) advisories. On the held-out test, CLAWAUDIT raises Semgrep recall from 21.7% (Pro baseline) to 66.8%, and CodeQL recall from 13.8% (security-extended) to 75.1%. Train/test gaps remain within 4 percentage points for all four configurations, indicating that the rules generalize to vulnerabilities unseen during rule writing. A preliminary live-code audit shows that these recall-oriented rules require manual triage, motivating semantic filtering before production deployment.

Training-free detectors of AI-generated images promise generator-agnostic deployment without classifier training, yet their reported numbers are rarely compared under a single controlled protocol. We audit two representative training-free scores -- an autoencoder-reconstruction score (AEROBLADE-style) and a noise-perturbation feature-similarity score (RIGID-style) -- plus a naive feature-kNN control, on a common 1,500-image GenImage-derived benchmark spanning seven generators and JPEG compression at quality 70 and 50. The audit yields three cautionary findings. (i) Implementation details masquerade as method differences: replacing the LPIPS backbone (AlexNet -> VGG-16) changes overall AUROC by +0.085, and switching between resize-to-512 and native-resolution preprocessing flips per-generator conclusions by up to 0.38 AUROC. (ii) Score direction is not a property of the method but of its hyperparameters: the RIGID-style score is inverted (AUROC < 0.5) on SD1.5 and Wukong at noise level sigma=0.05, recovers to >0.5 for every generator at sigma=0.01, and collapses to 0.15 at sigma=0.3. (iii) Dataset format bias inflates robustness claims: without unified re-encoding, AUROC under JPEG-50 exceeds the clean condition for the AlexNet-backbone reconstruction score; after bias correction the residual anomaly localizes to a single generator (BigGAN). The audited scores have complementary per-generator failure sets, but naive z-score fusion does not beat the best single score, indicating that exploiting complementarity requires direction-aware combination.

Financial chart question answering in regulated settings demands more than accuracy: practitioners must know which answers to trust before acting on them, and many institutions cannot send client data to external model providers. Yet existing chart-QA agents are accuracy-focused and opaque, and most assume proprietary API access; to our knowledge, none combines auditability with on-premise deployability without significant accuracy compromise. We present AgentFinVQA, a multi-agent pipeline that decomposes each query into planning, OCR, legend grounding, visual inspection, and verification, recording every step in a traceable Model Evaluation Packet (MEP) per sample. On FinMME, AgentFinVQA improves $+7.68$ pp over a primary-backbone matched zero-shot baseline with a proprietary backbone (Gemini-3 Flash; 71.24% vs. 63.56%, McNemar $p \approx 1.1 \times 10^{-16}$), and $+4.84$ pp with open-weights Qwen3.6-27B-FP8 served locally. The verifier's verdict also serves as a useful confidence signal (68.2% vs. 55.6% exact accuracy on confirmed vs. revised answers), enabling human-in-the-loop review routing. Error analysis shows that question misunderstanding, legend confusion and extraction error account for nearly two-thirds of failures and are the categories least detected by the verifier, identifying clear directions for future work. Together these results show that auditable, on-premise financial chart QA is practical and that the open-weights system keeps most of the accuracy gains while enabling full data residency. We release our code to support reproducible evaluation.

Machine unlearning for large language models (LLMs) aims to remove specified knowledge while preserving the rest of the model's capabilities. However, the boundary between knowledge to forget and knowledge to retain is often unclear, since related and even distant information may be entangled in the model. In this paper, we study LLM unlearning from a data-centric perspective and measure how unlearning effects propagate from the forget set to same-domain and distant-domain knowledge not after but before unlearning. We find a consistent decay pattern: collateral damage is strongest near the forget set, weakens with semantic distance, but does not disappear at domain boundaries. We further ask whether such damage can be audited before unlearning is executed. We formulate forget-set auditing as a pre-unlearning prediction task and analyze which data features are most predictive of downstream damage. Our results show that interaction features between the forget set and evaluation set provide the strongest signals, suggesting that collateral damage is partly reflected in data geometry before model updates occur. These findings position forget-set auditing as an early warning tool for identifying risky unlearning runs and designing more reliable unlearning procedures. Code and data are available at https://github.com/BartSu/PreUnlearn.

Tool-using language-model agents introduce security failures that go beyond unsafe text: they can disclose protected objects, write persistent memory, send messages, modify databases, or trigger harmful code and tool effects. Existing evaluations often collapse these stages into a single attack success rate, making it difficult to tell whether a model merely agreed with an attacker or actually produced observable harm. We introduce SafeClawBench, a staged benchmark for tool-using agent security with 600 controlled adversarial tasks across six attack families: direct and indirect prompt injection, tool-return injection, memory poisoning, memory extraction, and ambiguity-driven unsafe inference. SafeClawBench reports three separate endpoints: semantic attack acceptance, audit-visible harm evidence, and sandbox-observed tool/state harm. Evaluating five agent endpoints under four prompt-level policies, we find that these endpoints capture different failure modes. Without additional prompt protection, semantic failure rates vary widely across models, from 9.0% to 44.2%. Audited harm evidence is narrower than semantic failure, and under a separate executable protocol some matched task identities produce sandbox harm despite passing the Semantic Core call: in a 12,000-row matched analysis, 291 of 347 observed sandbox harms occur in rows that pass the semantic check. Prompt policies change endpoint outcomes, but their effects depend on both model and protocol. SafeClawBench provides a reproducible framework for comparing agent models and prompt-policy conditions without conflating textual compliance, evidence-supported harm, and executable state changes. The open-source dataset is available at https://huggingface.co/datasets/sairights/safeclawbench.

Reproducing research results from papers and released code is central to scientific progress. Existing works have introduced benchmarks to evaluate whether LLM agents can assist with reproducibility, but they are difficult to scale due to their reliance on substantial manual effort for data curation and evaluation. We introduce ReproRepo, a scalable framework for reproducibility evaluation that leverages human-raised GitHub issues as naturally occurring supervision on realistic reproduction blockers. We instantiate ReproRepo on 1,149 recent machine learning papers from major conferences and evaluate four frontier model-agent configurations. Our results show that LLM agents, even without executing code, can identify many real-world reproducibility problems from paper-repository pairs: the best agent in our study, namely Codex with GPT-5.5, surfaces at least one semantically related human-reported blocker for $\sim$90% of papers in the study. Further analysis shows that agents are particularly effective for surfacing visible failures and identifying the right semantic region, but may still be insufficient in exact localization. ReproRepo can serve as a reusable, scalable framework for future evaluations of LLM agents on real-world reproducibility auditing. Our code is released at https://github.com/LithiumDA/ReproRepo.

AI systems deployed in legal workflows hallucinate at rates that aggregate metrics report at ~52%, but this average conceals where errors concentrate and in which direction they run, leaving compliance officers without an actionable signal for trustworthy deployment. We present LegalHalluLens, an auditing framework with three components: typed hallucination profiles across four legally-motivated claim categories (numeric, temporal, obligation/entitlement, factual) over CUAD (Hendrycks et al., 2021); a Risk Direction Index (RDI) that reduces omission-versus-invention bias to a single deployment-comparable scalar; and a typed debate pipeline calibrated to both magnitudes and directions. Across 510 contracts and 249,252 clause-level instances we measure a within-model gap of approximately 38-40 pp between obligation/numeric and temporal claims that aggregate reporting hides, and show that two systems with matched 52% rates can carry opposite RDIs. The debate pipeline reduces fabricated detections by 45% with per-category gains tracking the diagnosis, matching commercial APIs with a substantially smaller backbone (4B active parameters). Typed profiles and RDI surface failure modes that aggregate metrics hide; we further show these diagnostics serve as calibration inputs for multi-agent debate pipelines, where Skeptic challenges and asymmetric gates targeted at measured failure modes outperform generically-tuned debate. The framework supports direction-aware procurement, accountability, and agent design for legal AI deployed in the wild.

Contrastive Language-Image Pre-training models are widely reused across downstream interfaces, including feature extraction, retrieval, reranking, and selection. Existing CLIP backdoor, however, usually validate attacks on a small attack-native task, leaving unclear whether the same poisoned checkpoint remains exposed, weakens, or becomes not applicable when reused through other interfaces. We introduce DIFE, a Deployment-Interface Footprint Evaluation framework that audits backdoored CLIP checkpoints across deployment interfaces. DIFE makes various evaluations comparable by specifying each interface's component readout, trigger channel, target event, reference condition, and metric. DIFE also introduces effective-footprint diagnosis to identify the reusable CLIP component or component combination that carries exposure and explains where risk transfers. Auditing reproduced CLIP backdoors with DIFE reveals a structured landscape: native success is not a checkpoint-level risk certificate, exposure follows component footprints, text-side poisoning does not yield textual-encoder control, and some coupled attacks remain mechanism-bound. This audit reveals a import gapin existing CLIP backdoors: a textual encoder that itself becomes a reusable carrier of adversarial behavior. We therefore introduce BadTextTower to fill this gap. BadTextTower produces strong text-conditioned retrieval, reranking, and selection exposure while leaving visual-only reuse nearly clean.

Medical vision-language models report strong chest radiograph accuracy, and this is increasingly read as evidence that they use the image. That inference is unsafe: a model exploiting finding-name priors scores like one that reads the scan, and no standard benchmark separates them. We introduce a causal audit that intervenes on the image, occluding the relevant region, occluding an irrelevant one, and swapping in another patient's same-label scan, and combines three behavioral metrics to test whether a correct answer depends on the image. Across nine systems, a text-only model with no image access reaches within 5.7 accuracy points of the best multimodal one, and a 119-billion-parameter multimodal model is statistically indistinguishable from a 7-billion text-only baseline. The audit splits the cohort into three models that ignore the image, one that is unstable, and five that use it selectively, for a subset of findings; the categories hold across a second dataset, resolution, and prompt phrasing. Against board-certified radiologists, a text-only model is statistically indistinguishable from a radiologist's accuracy while grounding at zero, whereas the image-using models ground at radiologist-comparable rates. Reported confidence flags ungrounded answers only when a model uses the image. Grounding audits, not accuracy, should gate clinical deployment.

Public AI evaluations are often read as terminal leaderboards, yet the underlying evidence is a selective time series shaped by reporting rules, benchmark revisions, and missingness. Repeated public archives for LiveBench and Open LLM Leaderboard v2 serve as the primary longitudinal record; LMArena provides a preference stress test; and GAIA and tau-bench contribute limited agentic pilots. Together, these archives instantiate a Bayesian inference problem: under a fixed reporting convention, one constructed terminal-only example over $1{,}000$ systems is compatible with two pre-terminal histories, yielding times of $23.03$ or $75.13$ to reach within $0.05$ of the ceiling under the same terminal-tail model. In synthetic posterior comparisons, action-facing diagnostics differ across observation regimes. The candidate selection-aware frontier model fails synthetic recovery, objective-archive prediction, preference transfer, and uncertainty calibration; correspondingly, fixed audit gates reject its stronger claims. An archive-and-adjudication protocol reconstructs public evaluation histories, isolates a verified timing boundary, and falsifies unsupported frontier claims.

The rapid adoption of generative AI and Large Language Models (LLMs) has spurred interest in synthetic data as a privacy-preserving alternative to sensitive real-world datasets. However, generating high-utility synthetic data often carries the risk of memorizing and regurgitating private information from the training corpus. In this work, we present a customizable empirical auditing framework designed to detect and explain such data disclosures. Our framework introduces a mechanism to distinguish between "true disclosures"-where the system directly reproduces a user's information-and "phantom disclosures''-where the system incidentally generates a user's data. By partitioning input data into training and holdout sets and applying rigorous statistical hypothesis testing, we determine if observed disclosures are consistent with strict privacy baselines, such as zero-learning or specific Differential Privacy (DP) bounds. Crucially, this approach requires no model access, no canary insertion, and no reference model training -only the synthetic output and a held-out control set. We demonstrate that this framework effectively functions as a membership inference attack, providing empirical lower bounds on privacy leakage that are tighter than prior data-based auditing methods. Our approach is model-agnostic, applies to any synthetic data generation mechanism, and requires orders of magnitude fewer computational resources than shadow-model or canary-based alternatives.

As large language models (LLMs) are increasingly deployed in user-facing systems, black-box jailbreak defense has become an important practical problem. Existing defenses often rely on known-attack coverage, prompt-level semantic judgment, or local runtime control, yet these paths can become unstable under evolving prompt packaging, expression rewriting, and structure manipulation. We observe that many black-box jailbreaks do not remove the harmful goal, but reorganize the information needed to express and execute it, thereby evading safety alignment while remaining recoverable during generation. Motivated by this observation, we propose DoubtProbe, a dual-branch inference-time defense framework that combines structural verification with semantic auditing and formulates black-box jailbreak defense as consistency checking under controlled transformation. The structural branch extracts a structured representation from the original request, reconstructs the request under representation constraints, and detects information-preservation failures between the original and reconstructed requests; the semantic branch audits the original prompt directly. We evaluate DoubtProbe against representative black-box defenses on jailbreak and benign-request benchmarks, and further test backbone transfer from Qwen2.5-72B to Llama-3.1-70B. Results show that DoubtProbe achieves a stronger and more stable defense-utility trade-off: on Qwen2.5-72B, it reduces the JBB attack success rate from 0.293 to 0.100 and the CodeAttack attack success rate from 0.152 to 0.001, while maintaining false positive rates of 0.022 and 0.016 on AlpacaEval and OR-Bench; the same pattern remains stable on Llama-3.1-70B. These findings show that structural inconsistency signals provide a practical and generalizable basis for black-box jailbreak defense, especially when combined with semantic auditing.

Travelers increasingly ask large language model (LLM) assistants which hotel to book, making these systems gatekeepers of property visibility -- yet what moves their recommendations is undocumented. We conduct a pre-specified algorithm audit using a randomized choice-based conjoint: across personas, prompt templates, and twelve open-weight and proprietary models, assistants choose among five hotels whose guest rating, review volume and recency, management response, chain affiliation, price, eco-certification, and list position are independently randomized. We estimate the average marginal component effect of each signal on the probability of recommendation. Guest rating and price dominate (a top rating raises selection by 31.6 percentage points; a high price lowers it by 30.0), reproducing human valence-and-price primacy but over-weighting eco-certification and ignoring management response. List position -- a content-free artifact -- shifts recommendations causally, worth about $12 per night. Stated reasons track revealed weights imperfectly. The findings ground generative engine optimization and the accountability of AI infomediaries in causal evidence.

The worldwide surge of authoritarianism, combined with the increasing central role in users' everyday lives, raises the question of to what extent specific models exhibit or promote authoritarian attitudes and characteristics. We introduce AuAu, a comprehensive benchmark that aims to assess the risk of LLMs generating responses with authoritarian tendencies. This benchmark combines three evaluation approaches: (i) psychometric questions from an extensive pool of 15 human validated instruments; (ii) contextual behavior vignettes probing intended actions in concrete situations; and (iii) responses to realistic user prompts. Unlike prior work, AuAu evaluates not only a general closeness towards authoritarianism but also the established sub-concepts Authoritarian Aggression, Authoritarian Submission, and Conventionalism. Evaluating 17 models from China, the EU, Russia, and the USA, we find that all tested models exhibit substantial authoritarian response rates under the psychometric evaluation, though rates drop significantly in increasingly more realistic downstream task. We further find that an authoritarian system prompt easily manipulates 15 out of 17 models to promote increased authoritarianism. Our results underscore the need for continued, systematic auditing of LLM-based AI systems to detect and ultimately mitigate undesired authoritarian tendencies in generated output. Our code and data are available at: https://github.com/andreaseinwiller/AuAu

Machine unlearning has been extensively studied in response to growing privacy concerns and regulatory requirements. However, auditing whether unlearning algorithms have truly erased the influence of specific data remains an open challenge. The lack of reliable and practical auditing mechanisms can lead to critical privacy risks, such as residual information leakage. This paper initiates a systematic investigation into whether existing unlearning algorithms can truly forget the designated data. We propose the first practical and general-purpose auditing framework for machine unlearning, inspired by the concept of proof of ignorance. Our framework addresses the key practicality limitations of existing methods by eliminating the need for retraining-from-scratch baselines, avoiding the training of large numbers of shadow models, and requiring no intrusive intervention in the original training process. To evaluate the effectiveness of our framework, we first conduct validation experiments to verify its soundness and completeness. We then perform comprehensive experiments across six datasets and ten representative unlearning methods. The results demonstrate that our framework reliably distinguishes between successful and failed unlearning. In particular, we observe that retraining-based and fine-tuning-based methods can achieve effective unlearning, even when the target data remain in the original dataset. In contrast, de-optimization-based methods fail to achieve true unlearning and instead degrade the model's performance. Fisher/Hessian-based methods also fail to unlearn requested data, even formal certification is provided. Moreover, we show that our framework is robust against fake unlearning attempts and generalizes well to large language models.

AI systems are increasingly used to assist consequential decisions in regulated domains such as auditing, finance, and healthcare. This creates a recurring obligation: an organization must be able to reconstruct, after the fact, which evidence informed a given conclusion, and to show that the record of that reasoning was not altered. Existing tools address related but distinct problems - model observability, drift monitoring, governance reporting - and are built for the machine-learning engineer operating a system, not the reviewer who must trace one specific conclusion back to its supporting evidence. We present AuditWeave, a lightweight Python library, with no runtime dependencies, that records the steps of AI-assisted and data-transformation workflows into a single append-only, hash-chained ledger. A small, system-agnostic event vocabulary spans both retrieval-augmented generation (RAG) pipelines and tabular/lakehouse transformations, so a conclusion that draws on both can be traced end-to-end through one record. Within a sealed ledger, any modification, reordering, insertion, or deletion of events is detectable through chain verification. We describe the design and evaluate recording overhead, scalability, and tamper-detection correctness on the reference implementation. The integrity guarantees cost tens of microseconds per event, and, as the hash-chain construction implies, verification flagged every injected mutation across four mutation classes over 2,000 randomized trials.

Existing financial-NLP benchmarks mostly evaluate prepared artifacts such as filings, tables, or extracted values. Real accounting begins earlier: source documents must be reconciled into cited journal entries, aggregated into a balance sheet, and checked for contradictions. We introduce FinBalance, a multi-document accounting reconciliation benchmark built from source-document bundles across eight industries, three period types, and five difficulty levels. Human-authored business scenarios, accounting policies, tax/FX treatments, document schemas, distractors, and inconsistency templates are composed by a deterministic generator whose ledger produces journal entries,balance sheets, and 23 inconsistency-code labels. On a 710-record evaluation split, six contemporary LLMs reach at most 46% exact final-balance-sheet accuracy. Four models show a 26-41 pp gap between BS_exact, the model's reported balance sheet, and BS_recon, the balance sheet obtained by replaying its entries through our ledger. Models often recover numerically plausible entries but fail to bind them to supporting documents and aggregate them consistently. Citation-pressure prompting barely changes document-linking errors, while ledger-feedback ablations substantially improve reported balance sheets and expose inconsistency-detection trade-offs. Expert finance reviewers validate the benchmark design and labels.

Synthetic and distilled student data are increasingly used to enable privacy-conscious learning analytics, yet their suitability for decision-facing institutional support remains uncertain. In dropout support, generated data must preserve not only predictive utility or distributional resemblance, but also the financial-status evidence used to guide advising, payment-plan assistance, and scholarship-related decisions. Method: This study introduces CaP-Eval, a decision-facing causal-privacy audit workflow for evaluating generated student data under a fixed estimand, timing-aware adjustment design, estimator set, and empirical privacy-governance screen. The workflow compares original, distilled, adversarial synthetic, statistical synthetic, and DPGNet privacy-oriented generated data on predictive utility, treatment-effect fidelity, robustness to alternative estimators, and local training-record proximity. Results: DPGNet and distilled data preserved the original financial-status treatment-effect structure more reliably than the adversarial and Gaussian Copula baselines. DPGNet preserved full direction and rank agreement across epsilon levels; epsilon = 10 produced the smallest non-original IPW and DML deviations, while epsilon = 1 and epsilon = 5 amplified several financial-status contrasts. Distilled data remained highly faithful but retained the strongest local training-record proximity signal. TabularGNet preserved qualitative directions with moderate attenuation, and Gaussian Copula compressed effect magnitudes. Conclusions: Predictive utility, privacy orientation, empirical disclosure signals, and causal fidelity diverged; generated student data require joint audits of direction, magnitude, overlap, and release-governance risk before decision use.

Audit risk assessment increasingly benefits from combining heterogeneous evidence sources, yet existing approaches typically produce point predictions without quantifying how well different evidence streams agree. We propose UMAR (Uncertainty-Aware Multi-Agent Risk Assessment), a framework that employs three specialized agents: an MD&A Text Agent, a Financial Ratio Agent, and a CAM Agent, each producing independent risk scores with calibrated uncertainty estimates. An Uncertainty Aggregator based on Dempster-Shafer evidence theory fuses these scores while explicitly measuring inter-agent conflict. We evaluate UMAR on a U.S. dataset of 3,200 firm-year observations from SEC 10-K filings (2019-2023), with financial restatement as the target label. Experimental results show that UMAR achieves an AUROC of 0.782 and a PR-AUC of 0.341, outperforming logistic regression, XGBoost, FinBERT, and single-agent and dual-agent LLM baselines. UMAR attains the lowest expected calibration error (ECE = 0.052) among all methods and identifies evidence-conflict patterns that correlate with actual restatement risk, offering auditors potentially actionable and interpretable risk signals.

Behavioral audits of Large Language Models on moral prompts measure what the model says, not the internal computation producing it. We use Transluce, an AI-driven mechanistic-interpretability platform, to examine LLaMA 3.1-8B-Instruct on 54 moral prompts in four batteries: 17 dilemmas, policy, and meta-ethical questions (B1); 6 role-playing scenarios (B3); and a controlled trolley contrast varying the switching mechanism with people fixed (B4, 15 prompts) or identity attributes with mechanism fixed (B5, 16 prompts). Two complementary metric families, five cluster-level metrics and a six-metric neuron-level panel, converge on a Situational Anchor Effect: domain-specific representations dominate the top of the activation list across every battery. The model's ethics-labeled capacity stays essentially constant; its salience (rank, priority, top-of-list presence) is highly sensitive to the interpretive frame the prompt selects. The B4-vs-B5 contrast confirms the model attends to whichever surface feature varies: aggregate ethics metrics are indistinguishable, but the dominant non-ethics distractor mirrors the design. A multi-temperature audit identifies a candidate ethics neuron (L16/N3837) stable across temperatures; a cross-model behavioral proxy on two frontier models yields preliminary evidence of divergence in self-reported moral focus, consistent with an Alignment Wrapper in which RLHF re-orders surface text without removing underlying domain-first frames. We unify these as Frame-Conditioned Moral Computation: the prompt's surface vocabulary selects a feature manifold, and the moral conclusion is downstream of that selection. Behavioral alignment must be supplemented by Mechanistic Alignment: a research program asking whether ethics-related features can be shown causally privileged under controlled frame variation, not merely loud in the explanation.

Selective prediction with distribution-free risk control promises that, with confidence 1-delta over the calibration draw, the error rate of accepted inputs stays below a user budget alpha. We audit this promise on signal-domain detectors -- machine anomalous-sound detection (ASD) and AI-generated-image forensics -- for four calibration rules: uncertified empirical thresholding (NAIVE) and certified Hoeffding, Clopper-Pearson (CP), and betting (WSR) upper confidence bounds. We report three findings. (i) NAIVE thresholding, common in practice, exceeds its declared budget in 49-73% of synthetic trials (n=200 calibration points) and in up to 68% of real-data splits: a false sense of safety rather than a broken theorem, since the rule never had a certificate. (ii) Tightness matters: CP and WSR certify substantial coverage where Hoeffding certifies none, with zero observed budget overruns under exchangeable splits. (iii) Under grouped deployment (unseen machine types or generators), certified rules overrun in 9-30% of trials -- far above delta -- showing the failure lies in the broken exchangeability premise, not in the bounds; a conservative per-group threshold restores validity at a severe coverage cost.

We consider the problem of uncertainty quantification for a pretrained classification model deployed under unknown distribution shift. We propose Audited Conformal Prediction (ACP), a method that leverages a small labeled dataset from the target population to train an auxiliary audit model identifying inputs where the legacy model is likely to fail. By integrating the audit model's outputs into the conformal prediction framework, ACP produces prediction sets that guarantee marginal coverage while achieving substantially higher conditional coverage in practice than existing approaches. We develop and analyze two complementary integration strategies -- one targeting marginal coverage with improved conditional performance, the other providing explicit group-conditional coverage guarantees -- and establish theoretical guarantees for both. Experiments on synthetic and real-world datasets validate the method and illustrate trade-offs between prediction set size and conditional coverage.

The "First Proof" benchmark [1] posed ten research-level mathematics questions to the strongest publicly available LLMs and found them consistently wrong-not silent, but confidently, fluently wrong. This paper asks why. Working from the per-question post-mortems in First Proof's Appendix A, I identify four failure modes: citation fabrication (F1), premise smuggling (F2), silent problem reformulation (F3), and local-to-global compatibility gaps (F4). I then audit eight one-shot proofs generated by Gemini 2.5 Flash on Questions 1, 2, and 5 of the benchmark, using two instruments built specifically to surface F1 and F2. The central finding is uncomfortable for anyone who sees retrieval-augmented generation (RAG) as the obvious fix: not one of the eight proofs contained a confirmed fabricated citation, yet every single one contained at least one load-bearing claim asserted as a "fundamental result" or "standard argument" with no justification attached. That failure mode-F2, premise smuggling-is invisible to citation verification by design. A premise-audit instrument I introduce flags it at 100% precision (5/5 judge-confirmed flags are true positives) and 50% proof-level recall in this corpus. The taxonomy and the audit together suggest that the right long-term objective is building inference-time pipelines that prevent these failure modes from occurring, not just detecting them after the fact. Index Terms--Large language models, mathematical reasoning, hallucination, premise smuggling, failure-mode taxonomy.

The removal of learned data from Machine Learning models through Machine Unlearning (MU) has been widely studied; however, there has yet to be an agreed-upon scheme for auditing MU. Existing work has shown that a dishonest model owner can falsify evidence to avoid executing MU, while curious auditors (and adversaries) can infer the privacy-sensitive properties of the model and its training data even with limited access. Yet auditing of MU under mutual distrust between the model owner and the auditor remains unexplored. We provide an information-theoretic proof for this scenario: for convex ML models, a generic audit scheme that relies solely on querying the model for \textit{behavioral} signals cannot identify insufficiently unlearned models without revealing membership information of the retained set. Therefore, auditing MU under the assumption of a dishonest model owner and an honest-but-curious auditor faces an inherent privacy-audit tradeoff. Our empirical results on convex models strongly supports this result, while further experiments demonstrate that this privacy-audit tension persists in non-convex models. Our results call for a more careful consideration of the privacy-audit tension under a realistic auditor threat model, and serve as a foundation for more scrutiny of designs of privacy-preserving audit schemes for the MU pipeline. We also release our code implementation at https://github.com/LiouTang/Behavioral-Unlearn-Audit.

We present a recipe-controlled decoder audit (RCDA) for structural transductive knowledge-graph completion (KGC). The audit asks a simple reporting question: before attributing gains to an encoder or training recipe, what changes when the decoder is swapped under the same recipe? Using ComplEx and DistMult as the primary controlled pair, with targeted RotatE/TransE spot-checks, we evaluate seven benchmarks. On five standard KGs, ComplEx-vs-DistMult differences are modest but consistent under our recipe (+0.005 to +0.012 MRR), whereas CompGCN-style encoder effects vary more by dataset. On small KGs, decoder effects become the main diagnostic: Kinship shows a stable ComplEx advantage of +0.143 MRR (6 seeds), while UMLS favours ComplEx by +0.022 MRR in a clean 6-seed server rerun but reverses in an earlier provenance variant. We therefore treat small-KG decoder choice as recipe- and provenance-sensitive rather than as a fixed dataset winner. We further show that decoder choice interacts with encoder depth on WN18RR, and that under our recipe L=0 ComplEx on YAGO3-10 reaches 0.6971 +/- 0.0048 MRR at d=128. The result is a compact audit protocol: report matched decoder rows, log small-KG provenance, and sweep decoder x depth before making encoder-level claims.

Transformer-based clinical language models are increasingly integrated into high-stakes clinical decision support pipelines, yet the computational mechanisms through which demographic associations encoded in medical documentation propagate into model probability distributions remain empirically underspecified. We present a systematic computational audit of representational bias in ClinicalBERT (Alsentzer et al., 2019), a BERT-based model pretrained on MIMIC-III discharge summaries, employing two complementary probing methodologies: Log Probability Bias Analysis (LPBA), which quantifies demographic descriptor-induced shifts in masked token probability distributions across behavioral and evaluative semantic categories, and Masked Language Model-based analysis (MLM), which probes internal representational structure for demographic agency attribution encoding across 98 real clinical sentence templates and eight intersectional race-gender combinations. Corpus frequency analysis operationalizes the distinction between statistical disparity and bias amplification by benchmarking model outputs against empirical term frequencies in the MIMIC-III training corpus. Of 32 statistically significant findings, 65.6% contradict observed corpus distributions, rising to 80% for Black patients and 87.5% for agency attribution under MLM probing, providing direct empirical evidence that representational bias in ClinicalBERT operates predominantly through model-internal amplification rather than training data inheritance. Keywords: natural language processing, clinical documentation, algorithmic auditing, representational bias, health equity 1

End-to-end (E2E) autonomous-driving planners trained by imitation are prone to statistical shortcuts: they associate scene elements that merely co-occur with expert actions (a roadside object, a building facade) with driving decisions, rather than the variables that causally determine them. Such causal confusion silently compromises reliability in long-tail scenarios, and it is difficult to detect, because prevailing open-loop metrics (L2 displacement and collision rate) are dominated by ego status and do not indicate whether a planner depends on spurious cues. Existing remedies based on causal-intervention training require retraining large models and cannot audit a planner that is already deployed. We present CADET, a training-free framework that audits, benchmarks, and repairs spurious reliance in pretrained E2E planners without any parameter update.

Semantic 3D occupancy provides a voxelized world state for autonomous driving and robot decision making, but object and rare-class errors can affect free-space interpretation, collision checking, and temporal state propagation. We show that a common VLM strategy, aligning 3D voxel or object features with crop-caption embeddings, improves text-space similarity without reliably improving closed-set occupancy mIoU. Motivated by this mismatch, we propose VISA, a training-time semantic auditing approach for existing occupancy world models. VISA queries an offline VLM on a representative crop of each physical object instance, obtains a structured audit with class hypotheses, plausible confusions, reliability, attributes, and evidence, and propagates it along the object track. The audit is grounded to matched 3D object voxels and distilled into semantic logits through reliability-weighted taxonomy, attribute-factor, and scene-level audit graph losses, while inference remains unchanged and requires no VLM. On nuScenes, averaged across three runs, VISA improves OccWorld from 19.06 to 20.05 mIoU and GaussianWorld from 21.36 to 21.91 mIoU; on GaussianWorld, object mIoU improves from 18.18 to 19.16 and rare-class mIoU from 15.60 to 16.79. These results suggest that VLMs are better suited to closed-set occupancy as reliability-aware semantic auditors than as generic caption-embedding targets.

Large Language Models (LLMs) are rapidly evolving into agentic systems that interact with external tools and environments, introducing new security risks such as indirect prompt injection attacks through untrusted external sources. Existing defenses mainly focus on blocking malicious content at inference time, and current red-teaming methods primarily optimize attack success. As a result, developers have limited visibility into how latent prompt injections emerge and propagate through agents. We propose PI-Hunter, an automated agentic auditing framework for proactive vulnerability exposure in LLM agents. PI-Hunter constructs realistic source-aware test cases and iteratively evolves them through feedback-driven exploration to induce agents to retrieve and reveal latent malicious instructions embedded within external environments. Extensive experiments across multiple benchmarks, agent architectures, attacks, and defenses demonstrate that PI-Hunter substantially improves vulnerability exposure and attack-surface coverage over strong automated red-teaming baselines, while remaining effective under existing prompt injection defenses.

Continual learning (CL) models often use experience replay to reduce catastrophic forgetting, but their robustness to replay sampling interference remains underexplored. Existing CL attacks alter inputs or training pipelines (poisoning/backdoors) and rarely include explicit auditable constraints, limiting realism. Here, auditability means a monitor can verify compliance from sampler-visible telemetry - e.g., logged replay index/label statistics - by checking that the realized replay class histogram stays close to a nominal baseline and that replay rate is unchanged per batch and/or over a rolling window. We study a limited-privilege insider who controls only replay index selection, not pixels, labels, or model parameters, while staying within auditable limits such as queue priorities. We introduce Amnesia, a replay composition attack that maximizes degradation under two budgets: a visibility budget delta bounding the TV/KL divergence from a nominal class histogram p0, and a mass budget f fixing the replay rate. Amnesia has two steps: (i) compute lightweight class utilities, such as EMA loss or confidence, to tilt p0 toward harmful classes; and (ii) project the tilt back into the delta-ball using efficient KL (exponential tilt) or TV (balanced mass redistribution) optimizers. A windowed scheduler enforces rolling audits. Across challenging CL benchmarks and strong replay baselines, Amnesia consistently lowers final accuracy (ACC) and worsens backward transfer (-BWT). The KL variant delivers high impact while remaining largely undetected under multiple audit schemes, including per-batch and rolling-window checks. The TV variant is more damaging but easier to detect, especially under tight per-class constraints. These results expose index-only replay control as a practical, auditable threat surface in CL systems and establish a principled impact-visibility trade-off.

Modern LLM training pipelines increasingly rely on other models to generate data, filter corpora, judge outputs, and guide development decisions. These dependencies are recursive: a model may depend on an upstream artifact whose own dependencies are documented only in separate releases and artifacts. As a result, the full dependency structure is fragmented across heterogeneous public artifacts, with complexity and recursive depth far outpacing humans' ability to trace. We introduce ModSleuth, an agentic system that recursively reconstructs LLM dependency graphs from public artifacts with source-grounded evidence. We find that the primary challenge is no longer information extraction, but defining what constitutes a dependency and reconciling artifact references across inconsistent documentation. We address these challenges through a formalization that distinguishes direct and indirect dependencies, represents heterogeneous pipeline roles through operation-centered relationships, and resolves artifact identities across names, versions, and repositories. Applying ModSleuth to four public-artifact-rich LLM releases, we recover 1,060 source-verified dependencies and construct large-scale dependency graphs of modern LLM development. These graphs reveal multi-hop license obligations, train-evaluation coupling, discrepancies between released and training-time artifacts, and documentation inconsistencies that would otherwise be difficult to uncover. We release ModSleuth and the resulting dependency graphs to support transparent analysis of the increasingly complex ecosystems underlying modern LLMs.

Research on bias in large language models (LLMs) has predominantly focused on third-person audits, which study how models represent or evaluate demographic groups as external subjects. However, this paradigm overlooks a structural blind spot because the user is absent from the audit. In practice, LLMs are used in open-ended, personal interactions, during which the model implicitly represents the user and adjusts its responses accordingly. When identical requests yield different responses depending on who is asking, bias manifests not in how the model describes others but in how it treats its interlocutor. We propose Situated Interaction Auditing (SIA), a user-centered framework for studying how user profile signals -- implicit sociodemographic markers, writing style, and stated identity -- systematically shape LLM response quality, content, and tone. We demonstrate the framework through a case study that intersects gender and socioeconomic status signals across multiple task domains and outline a research agenda for SIA as a new mission for natural language processing.

Agent skills let LLM agents reuse instructions, resources, tools, and workflows, but they also create a new place for malicious behavior to hide. A skill may look benign in its documentation or code while becoming harmful only when it is invoked with particular user requests, local assets, persistent state, or multi-step tool interactions. This makes purely static vetting brittle. We present Runtime Skill Audit (RSA), a dynamic analysis method that audits skills by asking what the skill-mediated agent actually does under targeted runtime conditions. Instead of testing every skill with the same generic tasks, RSA profiles risk-relevant interfaces, prepares the execution context needed to exercise them, and assigns security labels from the resulting trace evidence. We instantiate RSA on OpenClaw and evaluate it on 100 skills against representative static baselines. RSA achieves 90.0% accuracy with an 88.0% true positive rate and an 8.0% false positive rate, improving accuracy by 13.0 percentage points over the best static baseline. Under self-evolving attacks, static detectors collapse after one or two rounds, while RSA continues to detect 19--20 out of 20 malicious skills across rounds.

Compression progress is a long-standing proposal for intrinsic motivation: reward an agent when its world model becomes better at predicting or compressing experience. The folk claim is that this reward is "credible" because it is paid only for learning. We make this precise and prove it. If intrinsic reward is the signed decrease of a fixed sealed-audit loss, r_t = E(theta_{t-1}) - E(theta_t), then cumulative reward telescopes exactly to endpoint audit improvement, so no policy can push reward up indefinitely while true audit performance stagnates or degrades. For finite audit panels the same result holds with a sharp false-positive budget: cumulative empirical reward is at most true audit improvement plus 2 Delta_n(F, delta), the uniform audit deviation of the model class. This is horizon-free: adaptivity over time costs nothing once the sealed panel uniformly controls the class. The theorem also identifies the failure modes: the guarantee disappears if progress is clipped, scored on the agent's own stream, exposed to a high-capacity model on a reusable panel, or applied to a neural class that makes Delta_n vacuous. We give a Lean 4 mechanization of the structural core (telescoping, the finite-audit bound, finite Gibbs, and the entropy floor) and an experiment suite on ARC-TGI grid-transformation generators with adaptive holdout attacks. Experiments confirm the theory: finite-audit deviation scales as n^{-0.527}; signed progress resists clip-farming, stream leakage, and noisy-TV curiosity; naive reusable audits are exploitable by black-box scalar feedback, while standard release defenses keep the attack below the 2 Delta_n threshold. Signed compression progress on a sealed audit is an accounting signal of genuine improvement.

Claims about the robustness and fairness of deepfake speech detectors are only as credible as the datasets used to train and evaluate those systems. We present a dataset-level audit of the deepfake speech landscape. We compile and analyze 39 deepfake speech datasets, examining key attributes including accessibility, documentation, demographic and language coverage, dataset scale, and the underlying bona fide speech sources. Our audit reveals two important takeaways. Firstly, fairness assessment is largely infeasible because most datasets lack demographic metadata, and only a few contain gender or language labels. This prevents any meaningful subgroup analysis and leaves other demographic attributes unaddressed. Secondly, we identify substantial overlap in underlying bona fide source corpora across datasets, which can undermine cross-dataset evaluation and lead to overstated generalization claims.

Interpretability methods routinely use population-level summary statistics over observed model behaviour to license claims about the effects of targeted interventions on specific computations; in Pearl's terms, they treat rung-1 associational evidence as if it supported rung-2 interventional conclusions, a move whose validity is rarely tested. We examine one concrete instance: the use of routing statistics in Mixture-of-Experts (MoE) pruning, where utilization rates, activation norms, and routing weight distributions are treated as predictors of which experts can be removed without functional cost. A token-level interventional audit across three high-redundancy MoE architectures (OLMoE-1B-7B-0924, Qwen1.5-MoE-A2.7B, DeepSeek-V2-Lite) finds no observational metric predicts causal expert importance in any model: across all 60 metric-layer combinations effect sizes stay below Cohen's $d = 0.23$, and no metric is reliably positive under our corrected, dual-test criterion. A per-token routing weight control, run with identical $n$, rules out insufficient power, recovering a signal whose CI excludes zero at OLMoE's final MoE layer ($d = +0.231$, 95% CI $[+0.09, +0.37]$, $p = 0.0013$). Existing pruning methods succeed in this regime not by identifying dispensable experts but because early-layer redundancy renders most selection criteria interchangeable. Our results provide an explicit counterexample to the common inferential step from population-level observational summaries to token-level interventional claims about expert importance, and illustrate how interventional audits can calibrate the evidential standards for interpretability claims.

Lipschitz-style individual fairness formalizes the idea that semantically similar examples should receive similar predictions, but its evaluation in multi-task learning (MTL) can be confounded by method-induced representation scales. This paper identifies threshold confounding: when the auditing tolerance is derived from each model's own representation distances, different algorithms are compared under different semantic thresholds. A threshold-drift analysis further shows how Bias rankings can change and identifies sufficient conditions for ranking preservation. We propose \textbf{ReLiF}, a reliability-aware framework that separates evaluation-time fixed-$δ$ auditing from training-time controlled regularization. ReLiF uses a shared reference tolerance for comparable auditing and a violation-rate feedback controller to keep the Lipschitz surrogate active without letting it dominate stochastic training. This work also develops supporting analysis for threshold drift, reference-tolerance selection, and the relationship between the huberized training surrogate and its unsmoothed positive-margin counterpart. Experiments on clinical time-series benchmarks and NYUv2 (NYU Depth V2) dense prediction show that fixed-$δ$ auditing exposes utility--fairness trade-offs that method-dependent thresholds can obscure. On NYUv2 with a ResNet50 backbone, ReLiF achieves competitive utility while substantially reducing aligned bias under shared fixed thresholds. On clinical benchmarks, ReLiF yields controlled fairness-regularized trade-offs, while fixed-$δ$ auditing reveals that task-balancing baselines can sometimes achieve lower bias and that genuine utility--fairness trade-offs persist. These results support fixed-$δ$ auditing as a semantically consistent protocol for evaluating Lipschitz fairness in MTL.

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

This paper presents AuditBench, a new benchmark dataset for evaluating the capabilities of LLMs at investigating security-related system audit logs. We design and use this benchmark to explore the performance of LLMs on four log-investigation tasks that incident response teams commonly perform, ranging from triaging alerts generated by detectors to identifying persistence mechanisms on compromised systems. AuditBench consists of system audit logs collected from Linux and Windows machines, and spans over 50 different security investigation scenarios, including both malicious and benign activity. Using our benchmark, we evaluate and analyze the performance of five frontier LLMs at analyzing audit logs for attack investigations. Our analysis illuminates how LLM performance and error profiles vary according to different design choices, such as differences in model size, data representation, prompt construction, and specific investigation tasks. Additionally, we characterize the quality of the explanations produced by LLMs and the types of errors that models make across our benchmark. Collectively, our work provides a foundation for assessing the capabilities of LLMs for investigating security logs, novel insights for practitioners using LLMs in security operations, and important directions for future research.

Biomedical text mining, scientometrics, and the construction of training corpora for biomedical large language models (LLMs) all assume that the abstract text returned by a bibliographic API faithfully reproduces the published abstract. This pre-registered audit (OSF osf.io/269b5) tests that assumption for four widely used public APIs (PubMed E-utilities, Crossref, OpenAlex, Semantic Scholar) against PubMed Central (PMC) JATS XML as a common ground truth. From a complete enumeration of the PMC Open Access subset for 2024 (about 700,000 records), a simple random sample of 4,000 English-language research articles was drawn; for each, we recorded whether Unicode characters from four pre-specified classes present in the JATS abstract (typographic punctuation, mathematical/scientific symbols, Greek letters, special whitespace) were preserved by each API. Two systematic, deterministic losses met the pre-registered criterion (upper 95% CI bound below 5%): the PubMed AbstractText field preserved typographic punctuation in only 0.6% of eligible abstracts (95% CI 0.3-1.0%), and OpenAlex preserved special whitespace in 0% (0.0-0.4%). A blinded mechanism audit attributed the first loss to character substitution and the second to inverted-index serialization. Mathematical symbols and Greek letters were preserved faithfully (over 95%) by all four APIs. Separately, Crossref returned no abstract for 24.6% of papers (coverage 75.4%, 95% CI 74.1-76.7%), concentrated in specific publishers (Elsevier and ACS: 0%). Character-level fidelity is therefore API-dependent and undocumented: the same publisher-deposited JATS text carries different surface signatures depending on the serving API, with direct consequences for tokenization-sensitive bibliometrics, corpus construction, and character-level indicators of LLM-assisted writing.

Medical vision-language models (VLMs) are evaluated on public benchmarks whose images and question-answer pairs have been freely downloadable for years, yet reported accuracy assumes these examples were absent from pretraining. We audit open VLMs on SLAKE-En, PathVQA, VQA-RAD, and an auxiliary public OmniMedVQA mirror using four detector families: image-side near-neighbour overlap against PMC-OA-beta, canonical-order exchangeability, cohort-relative Min-K%++ tail enrichment, and cross-model top-K overlap. We find measurable image-side source overlap on SLAKE-En: 19.8% of images are flagged under SigLIP-B-16 and 4.2% under SigLIP-SO400M, while out-of-domain controls produce 0/2000 flags. Manual adjudication shows same-modality, same-projection matches to different patients rather than verified pixel-level duplicates, so we interpret this as source or distributional overlap rather than confirmed per-image memorization. On the text side, Qwen2.5-VL on SLAKE-En shows a canonical-order exchangeability signal that survives ordering ablation and external non-medical baselines. On the OmniMedVQA mirror, exchangeability fires for five medical and general VLMs while BLIP-2 remains clean. In contrast, cohort-relative Min-K%++ tail enrichment and cross-model top-K overlap collapse under an external pre-domain baseline: BLIP-2 reproduces the apparent positive signals despite lacking plausible medical-VQA exposure. We conclude that these cohort-relative detectors are unreliable as standalone membership-inference signals on small medical-VLM cohorts.

BCI-to-agent pipelines turn decoded neural activity into an authorization channel for tool-use agents, exposing a new attack surface we call \emph{brain-prompt injection}: signal-side perturbations, context-only injections, and adaptive dual-decoder attacks can all change the routed action while EEG-side or text-side monitors remain blind. Route safety in this stack depends on what the audit log can observe, not on decoder accuracy or agreement alone. We define a Route-Safety Audit Contract: a minimal log schema, denominator hierarchy, and endpoint specification, and prove an audit-schema separation theorem together with a C3 attacked-dependence decomposition; clean agreement and marginal robustness do not identify the joint term that controls C3 routing. As a calibration layer on top of the contract, we apply split-conformal calibration to a non-oracle EEG confirmation channel and report the resulting false-accept frontier under an explicit threat-archetype matrix. We instantiate the contract on EEGMMI native left/right command-control over 5{,}400 events, harmless tool stubs, and seed/case denominators. Provenance blocks C2 routes ($0.000$); agreement-plus-provenance routes C3 flips ($1.000$); confirmation-plus-provenance routes them ($0.000$). The conformal frontier reaches FAR $0.000$ at clean utility $0.150$ for $α=.005$ and FAR $0.119$ at clean utility $0.452$ for $α=.10$ under acquisition isolation; an attacker-controllable confirmation channel breaks the bound to $\approx\!1$. Subject-cluster bootstrap confirms these intervals on $60$ subjects; cross-architecture (TinyEEGNet, EEGNetV4) and capacity-sweep results show within-regime saturation. Mediation and confirmation reduce risk; they are not intent certificates.

Knowing how often a language model fails does not explain where its errors concentrate. When auditors examine many explanations, the strongest observed pattern may arise by chance. We introduce Janus, a procedure for checking proposed error patterns before reporting them. Janus starts with a fixed list of yes/no properties of the examples being evaluated, such as whether the input is long. For each property, it compares the model's error rates on examples with that property and those without it. To see how large a difference can arise by chance, it repeats this calculation after shuffling the yes/no labels across examples without changing the group sizes. These shuffled properties are called decoys. A pattern is reported only if the size of its error difference meets a threshold set using decoys. On separate held-out examples, the same group must still have the higher error rate and the difference must meet a minimum, which was chosen in advance. In a controlled experiment, where the model must find a code in documents containing tables of staff, projects, and renewal codes, Janus confirms five related patterns of higher error rates on tasks requiring more lookups across tables. It also confirms a sixth pattern: lower error rates on examples with the needed information at the ends of the tables. In our samples from the MuSiQue and LongBench v2 public benchmarks, SliceLine finds groups with high error rates, while Janus reports no confirmed error patterns for the example properties we chose to test. For comparison, we use standard tests that shuffle errors and account for testing many candidates. With the same holdout check, they confirm two to six controlled patterns, depending on the test and threshold, and none on either benchmark. In simulations with no real error patterns, Janus reports false patterns more often than Benjamini-Hochberg, depending on the decoy count.

We study the problem of auditing a black-box algorithmic decision-maker from observable inputs and outputs alone. Our main result is an exact decomposition: under precisely characterized conditions, the cumulative \emph{regret} of a dynamic policy equals the sum of per-period covariances between the cost vector and the policy's decision. This extends the single-period identity of Aldridge~(2026) to the full multi-period setting of stochastic dynamic programming. We prove the identity holds exactly under i.i.d. costs and mean-unbiased Markov policies, derive closed-form bias corrections for non-stationary and time-varying cases, and establish the discounted-horizon analog. A Bellman recursion for the covariance regret functional connects the result to standard reinforcement learning algorithms; for rolling-window policies, the estimation-error bias is $O(d/w)$. The decomposition has direct implications for algorithmic auditing in strategic environments: in platform mechanism design, it provides a welfare-based audit metric without access to the agent's private type; in repeated games, covariance reduction is a sufficient condition for policy improvement; in procurement and ad auctions, the bias correction quantifies welfare loss from strategic misreporting. The associated trajectory estimator is consistent, asymptotically normal with HAC variance, and computable in $O(T \cdot nd)$ time. This makes the proposed approach a tractable, model-free audit tool for platform mechanisms, algorithmic portfolio strategies, and any sequential decision system subject to external performance review.

Deep learning on physiological time series is interpreted through domain-specific features -- oscillatory rhythms in EEG, morphological complexes in ECG -- yet these signals sit atop a broadband aperiodic 1/f-like envelope that covaries with arousal, age, and pathology. We introduce a spectral audit framework combining aperiodic/periodic decomposition, phase-preserving Fourier interventions, sham controls, and simulation validation. Aperiodic reliance was task-dependent and architecture-general: across six neural architectures, flattening drops exceeded 0.42 balanced-accuracy points for sleep-wake classification, reached 0.07-0.13 for clinical abnormality detection, and remained minimal for motor imagery. Six of seven EEG foundation models showed FDR-significant aperiodic reliance on clinical EEG; age/sex and recording-era controls reduced but did not eliminate the effect. Applying the audit to PTB-XL ECG revealed neural drops of 0.32--0.36 persisting after demographic matching, confirming this confound class extends beyond EEG. Aperiodic controls should become standard for interpretable physiological time-series deep learning.

This paper reads six engine-level measurements together -- 1.1 host attack surface, 1.2 information leakage, 1.3 defense-in-depth stackability, 1.4 public CVE history, 1.5 patch cadence, and 1.6 upstream fuzzing posture -- to describe how five AI-sandbox products isolate guest code from the host kernel. No single axis is a sufficient basis for a comparative judgement; the cross-axis reading is the load-bearing analysis. Three high-level findings: (1) engine classes (microVM, userspace kernel, OCI container) separate cleanly on every architectural axis, but products within a class do not; (2) product pin policy is the dominant operator-facing variable -- engine-side patch latency aggregates to ~0 days for coordinated disclosures, while downstream lag spans 0 days to 471+ days to "opaque" to infinity; (3) fuzzing investment splits into three tiers, and the strongest combination -- microVM x continuous public fuzzer -- is unoccupied in this set, leaving the "0 published CVEs x no upstream fuzzer x no academic study" intersection structurally unmeasured. We report per-axis orderings, per-product portraits, and a threat-model qualification matrix; no overall ranking is proposed. Companion repository (code, Apache-2.0): https://github.com/orbitalab/RnD-ai-sandboxes-sec-study-part-1. License: CC BY 4.0.

Large language models (LLMs) are increasingly released and deployed through opaque development and deployment pipelines, enabling model providers to inject intentional, provider-specific policies without officially announcing them. As a result, various models have been reported to generate responses reflecting proprietary rules and organizational interests, leading to censorship or misinformation on controversial topics. However, systematic identification of such alignment remains a fundamental challenge, complicated by the ambiguity of what ``proprietary'' entails in different contexts. In this paper, we propose a statistical framework for detecting proprietary alignment in black-box language models via comparative behavioral analysis. Our approach quantifies systematic deviations between the responses of a target model and those of a reference set of baseline models in a shared semantic space. By evaluating relative behavioral divergence rather than absolute correctness, our framework enables principled auditing under black-box access. Applied to several widely discussed but previously unquantified cases, it provides a systematic and scalable basis for external assessment of provider-specific alignment behavior in large language models.

As large language models are increasingly deployed in high-stakes settings, there is a growing need for tools that audit not only model outputs but also the internal computations that produce them. Circuit analysis is a central approach in mechanistic interpretability, but it is typically target-conditioned, explaining a single prompt paired with a chosen completion. This target-conditioned setup can obscure heterogeneity across a model's continuation distribution. We introduce distribution-level unsupervised feature discovery, which clusters sampled continuations using both semantic content and sequence-level mechanistic attributions, without manually specifying target outputs. Our method represents each continuation with a semantic embedding and a prefix-to-continuation attribution signature, then optimizes a rate-distortion objective that trades off semantic coherence, mechanistic consistency, and cluster granularity. Across clustering and steering analyses, the discovered clusters expose continuation modes that single-view baselines miss and provide interventional evidence that cluster signatures correspond to actionable mechanistic factors. Overall, our approach complements circuit analysis and behavioral evaluation by providing a scalable audit of the mechanisms underlying a model's continuation distribution.

Large Language Model (LLM) safety has often been evaluated at the behavior level, which provides limited evidence of internal robustness, as these evaluations target outputs rather than representation-level vulnerability under intervention. We formalize this discrepancy as the audit gap: the difference between behavioral safety and robustness under intervention. To study this gap, we construct dissociated models that preserve safe outward behavior while remaining vulnerable in the latent space. We introduce an intervention-based evaluation framework to test model robustness through soft interventions in parameter and latent spaces, including harmful fine-tuning and layer-wise latent perturbations. To formalize the evaluation, we propose the Latent Vulnerability Score (LVS) to measure how easily harmful behavior can be elicited by bounded latent perturbations. Using this evaluation framework, we show that behavioral safety metrics are insufficient measures of representation-level robustness across multiple safely and unsafely aligned state-of-the-art models. Notably, dissociated models show substantially elevated LVSs despite comparable refusal behavior under harmful intervention, with intermediate representations being the most sensitive to intervention. Our results suggest that behavioral safety evaluation alone provides an incomplete picture of model robustness, motivating representation-aware audits of latent vulnerability and observable behavior.