LLM Auditing

LLM: Large Language Model

Follow this topic Ask about this topic

Momentum

59 papers in the last four weeks, up 79% on the four weeks before. 0.6% of all new papers.

Jul 13Week of Sep 28

Latest papers 453

May 18, 2026cs.AI

What Does the AI Doctor Value? Auditing Pluralism in the Clinical Ethics of Language Models

Medicine is inherently pluralistic. Principles such as autonomy, beneficence, nonmaleficence, and justice routinely conflict, and such ethical dilemmas often sharply divide reasonable physicians. Good clinical practice navigates these tensions in concert with each patient's values rather than imposing a single ethical stance. The ethical values that large language models bring to medical advice, however, have not been systematically examined. We present a framework for auditing value pluralism in medical AI, comprising a benchmark of clinician-verified dilemmas and an attribution method that recovers value priorities directly from decisions. The ecosystem of frontier models spans physician-level value heterogeneity, and models discuss competing values in their reasoning (Overton pluralism) before committing to a decision. However, individual model decisions are near-deterministic across repeated sampling and semantic variations, failing to reproduce the distributional pluralism of the physician panel. Across benchmark cases, these consistent decisions reflect committed, systematic value preferences. While most model priorities fall within the natural range of inter-physician variation, some significantly underweight patient autonomy. A single LLM deployed without regard for its value priorities could amplify those priorities at scale to every patient it serves. Without explicit efforts to balance ethical perspectives with one or multiple models, these tools risk replacing clinical pluralism with a deployment monoculture.

Payal Chandak, Victoria Alkin, David Wu +11

Moral Reasoning in Language Models Healthcare LLM Auditing Clinical Decision-Making Clinical Language Model Evaluation

May 18, 2026cs.CY

Position: Generative Engine Optimization Creates Underexamined Risks, Governance Must Target Concentration, Disclosure, and Academic Blind Spots

Large language model (LLM) answer engines are increasingly used for information seeking, shifting visibility from ranked lists to synthesized answers. This enables Generative Engine Optimization (GEO), which targets LLM answer engines' evidence pool and generation. We analyze the search engine optimization (SEO) to GEO transition to identify two risks: (i) concentrated influence from low contestability and system sensitivity, and (ii) undisclosed commercial influence embedded in evidence and reasoning. We then formalize a general GEO pipeline to locate where optimization acts and compare academic and industry practices, revealing a third risk: (iii) academic-industry blind spots driven by visibility and evaluation asymmetries between offline setups and deployed systems. This position argues the need for answer-level governance and measurement: stronger contestability, high-precision disclosure, black-box auditing of material influence, and deployment-aligned metrics for exposure persistence.

Yizhu Wen, Nan Zhang, Haohan Yuan +3

School of Informatics, Computing, and Engineering, Indiana University Bloomington, Bloomington, USA · The Media School, Indiana University Bloomington, Bloomington, USA · School of Data Science, University of North Carolina at Charlotte, Charlotte, USA +1

Generative Engine Optimization LLM Auditing AI Governance

May 17, 2026cs.LG

Auditing Reasoning-Trace Memorization Claims after Unlearning with Head-Conditioned Canaries

Evaluations of unlearning on reasoning models sometimes show a bypass pattern. The answer side looks unlearned, but the model's own thinking trace keeps emitting the forgotten content, and the gap is taken as evidence that the weights still remember. We audit this reading on DeepSeek-R1-Distill-Qwen-7B with LoRA-memorized fictional authors and NPO unlearning, conditioned on a six-token canary head. On one seed, swapping the thinking trace for a short non-canary prefill on the same weights drops the answer rate by as much as the bypass gap itself, whether the prefill mimics the training template or not. On a second seed the bypass gap shrinks rather than vanishing, and the prefill swap reverses direction and brings the answer rate to ceiling. A positive parser-split bypass gap thus does not by itself identify hidden weight-level memorization, and does not rule it out either. On a different distillate the same metric flips sign because the parser cannot find the closing tag. We recommend a decode-time template swap as a cheap sanity check alongside the canonical audit.

Yanhang Li, Zhichao Fan, Zexin Zhuang

Northeastern University, USA · University of Illinois Urbana-Champaign, USA · Southern Methodist University, USA

Memorization in Language Models LLM Auditing Neural Network Memorization Memorization in Generative Models LLM Unlearning Large Reasoning Models+2

May 16, 2026cs.LG

Representation Signatures and Risk-Feedback Alignment in LLM Trading Agents

We study behavioral alignment and representation dynamics of large language model (LLM) agents in financial decision environments. TradeArena, an auditable trading-agent testbed with risk reports, execution simulation, memory, and replayable trajectories, lets us analyze how rationales, positions, and interventions evolve under market stress. Code and data artifacts are available through the TradeArena repository. We find pre-failure signatures: planning embeddings drift from normal centroids, fused plan-risk representations separate normal from pre-drawdown states, and local manifolds exhibit effective-rank contraction. Across 80 rolling failure anchors and eight LLM trajectories, this pattern persists across hash, LSA, Transformer, and white-box hidden-state probes. Stress tests with CoT-free target weights, lexical controls, OHLCV noise, and false audits show that rationale-level contraction can vanish without rationales, while intent-space and fused signatures remain informative. Structured risk feedback can act as an external alignment signal without fine-tuning, but not as a universal performance enhancer: true audit feedback improves calibration for some models, returns for others, and exposes cases where placebo or hidden feedback has higher short-horizon return but weaker alignment diagnostics. A 51-stock intraday experiment reveals a correlation blind spot: LLM rationales justify exposure to coupled assets that the risk layer clips. Finally, a financial-audit task suite shifts comparison from ``which model trades best'' to whether models can audit trajectories, respect execution boundaries, reproduce artifacts, and avoid claim overreach. These results support a research claim, not a profitability claim: auditable risk feedback and representation trajectories reveal when LLM financial reasoning is aligning, drifting, or failing.

Weicheng Xue

Virginia Tech

LLM Alignment LLM Auditing LLM Interpretability LLM Agent Evaluation AI Agent Benchmarks LLM Agents+1

May 16, 2026cs.CL

GPF-LiveNews: A Streaming Evaluation Protocol for Group-Conditioned Framing in Large Language Models

Deployed language models are evaluated in a non-stationary environment: model versions, retrieval layers, safety systems, and real-world inputs all change over time. Static bias benchmarks remain useful, but they do not show how models frame newly emerging events for different prompted audiences. We introduce GPF-LIVENEWS, a streaming evaluation protocol and benchmark snapshot for auditing group-conditioned framing in open-ended LLM outputs. The protocol expands fresh BBC/Reuters news anchors across 42 identity labels and seven prompt families, then evaluates response bundles using semantic-sensitivity and sentiment-disparity signals. In a pilot over 12 monitoring runs and 23 hosted models, Policy/Action prompts produce the strongest semantic movement, while sentiment variation is flatter across dimensions and prompt families. The released artifact includes article metadata, prompt templates, instantiated prompts, model-output metadata, score tables, documentation, and reproduction scripts. We interpret all scores as observed-window audit signals for human review, not as permanent fairness rankings or direct proof of harmful bias.

Mohd Ariful Haque, Fahad Rahman, Kishor Datta Gupta +1

Clark Atlanta University · United International University

LLM Evaluation LLM Auditing Social Bias in Language Models Language Model Bias Evaluation

May 15, 2026cs.CY

AI-Mediated Communication Can Steer Collective Opinion

Generative artificial intelligence (AI) is increasingly integrated into the online platforms where humans exchange opinions; large language models (LLMs) now polish users' posts on LinkedIn and provide context for content shared on X. While prior work has shown that AI can express biased opinions and shape individuals' opinions during human-AI interactions, less attention has been paid to its influence on collective opinion formation when mediating human-to-human communication. We address this gap via a combination of empirical and theoretical analyses. We show empirically that LLMs from multiple popular families introduce directional biases when instructed to edit human-written texts on contested topics, for example, nudging texts in favor of gun control and against atheism. Building on this observation, we introduce a mathematical model of opinion dynamics in which an AI system sits between users on a social network, transforming the opinions they express and perceive. By analytically characterizing the equilibrium of this model and performing simulations on real social network data, we show that biases introduced by AI in human-to-human communication can be amplified through the network and shift collective opinion in their direction. In light of these findings, we investigate whether such biases are controllable by online platforms. We audit the "Explain this post" feature on X and find evidence of pro-life bias in Grok's outputs on abortion-related content, which we trace back to specific design choices. We conclude with a discussion of the broader implications of our findings in relation to ongoing legislative efforts in the European Union.

Stratis Tsirtsis, Kai Rawal, Chris Russell +2

1Hasso Plattner Institute · 2Oxford Internet Institute, University of Oxford · 3Weizenbaum Institute

Opinion Dynamics LLM Auditing Political Bias in Language Models Language Model Bias Evaluation

May 15, 2026cs.AI

Fully Open Meditron: An Auditable Pipeline for Clinical LLMs

Clinical decision support systems (CDSS) require scrutable, auditable pipelines that enable rigorous, reproducible validation. Yet current LLM-based CDSS remain largely opaque. Most "open" models are open-weight only, releasing parameters while withholding the data provenance, curation procedures, and generation pipelines that determine model behavior. Fully Open (FO) models, which expose the complete training stack end-to-end, do not currently exist in medicine. We introduce Fully Open Meditron, the first fully open pipeline for building LLM-CDSS, comprising a clinician-audited training corpus, a reproducible data construction and training framework, and a use-aligned evaluation protocol. The corpus unifies eight public medical QA datasets into a normalized conversational format and expands coverage with three clinician-vetted synthetic extensions: exam-style QA, guideline-grounded QA derived from 46,469 clinical practice guidelines, and clinical vignettes. The pipeline enforces system-wide decontamination, gold-label resampling of teacher generations, and end-to-end validation by a four-physician panel. We evaluate using an LLM-as-a-judge protocol over expert-written clinical vignettes, calibrated against 204 human raters. We apply the recipe to five FO base models (Apertus-70B/8B-Instruct, OLMo-2-32B-SFT, EuroLLM-22B/9B-Instruct). All MeditronFO variants are preferred over their bases. Apertus-70B-MeditronFO improves +6.6 points over its base (47.2% to 53.8%) on aggregate medical benchmarks, establishing a new FO SoTA. Gemma-3-27B-MeditronFO is preferred over MedGemma in 58.6% of LLM-as-a-judge comparisons and outperforms it on HealthBench (58% vs 55.9%). These results show that fully open pipelines can achieve state-of-the-art domain-specific performance without sacrificing auditability or reproducibility.

Xavier Theimer-Lienhard, Mushtaha El-Amin, Fay Elhassan +5

LiGHT, EPFL

Healthcare LLM Auditing Clinical Decision Support Medical LLMs LLM Training

May 15, 2026cs.AI

Formal Methods Meet LLMs: Auditing, Monitoring, and Intervention for Compliance of Advanced AI Systems

We examine one particular dimension of AI governance: how to monitor and audit AI-enabled products and services throughout the AI development lifecycle, from pre-deployment testing to post-deployment auditing. Combining principles from formal methods with SoTA machine learning, we propose techniques that enable AI-enabled product and service developers, as well as third party AI developers and evaluators, to perform offline auditing and online (runtime) monitoring of product-specific (temporally extended) behavioral constraints such as safety constraints, norms, rules and regulations with respect to black-box advanced AI systems, notably LLMs. We further provide practical techniques for predictive monitoring, such as sampling-based methods, and we introduce intervening monitors that act at runtime to preempt and potentially mitigate predicted violations. Experimental results show that by exploiting the formal syntax and semantics of Linear Temporal Logic (LTL), our proposed auditing and monitoring techniques are superior to LLM baseline methods in detecting violations of temporally extended behavioral constraints; with our approach, even small-model labelers match or exceed frontier LLM judges. Our predictive and intervening monitors significantly reduce the violation rates of LLM-based agents while largely preserving task performance. We further show through controlled experiments that LLMs' temporal reasoning shows a pronounced degradation in accuracy with increasing event distance, number of constraints, and number of propositions.

Parand A. Alamdari, Toryn Q. Klassen, Sheila A. McIlraith

University of Toronto, Vector Institute, Canada

LLM Auditing Formal Verification AI Agent Monitoring AI Governance AI Safety Linear Temporal Logic+1

May 15, 2026cs.LG

A Unified Perturbation Framework for Analyzing Leaderboard Stability and Manipulation

Evaluation leaderboards such as LMArena play a central role in benchmarking large language models by aggregating pairwise human preferences into model rankings, yet the robustness of these rankings remains poorly understood. We present a unified perturbation framework for analyzing Bradley-Terry leaderboards under structured data modifications using influence-based approximations. Our framework studies three match-level perturbations -- Drop, Add, and Flip -- together with player removal, and evaluates their effects on top-k membership, global ranking consistency via Kendall's tau, and confidence-interval-based uncertainty. Across Chatbot Arena and six additional pairwise-comparison datasets, we show that modern leaderboards are non-robust across all three objectives: sub-1% targeted perturbations can change the top-ranked model, degrade Kendall's tau, and alter confidence intervals. Beyond robustness auditing, we show that the same influence scores enable efficient targeted perturbations, promoting or demoting specific models and reducing target-model uncertainty with fewer actions than previous manipulation and active-sampling baselines. By summarizing these effects with normalized dataset-level robustness scores, our framework provides a practical and helpful tool for auditing leaderboard stability and motivating more robust evaluation protocols.

Hosna Oyarhoseini, Jimmy Lin, Amir-Hossein Karimi

University of Waterloo, Waterloo, Ontario, Canada.

LLM Evaluation Learning to Rank LLM Auditing Pairwise Ranking Bradley–Terry Model

May 15, 2026cs.CR

Asking Back: Interaction-Layer Antidistillation Watermarks

Detecting unauthorized knowledge distillation from a deployed LLM API is hard because the defender controls neither the attacker's training pipeline nor the next-token logits. Existing defenses operate on the teacher's output tokens -- biasing the next-token distribution (green-list watermarks, cryptographic schemes, antidistillation sampling) or rewriting outputs after generation. Recent work shows a paraphrasing attacker can strip these signals without losing the underlying knowledge. We propose interaction-layer antidistillation watermarks, which move the trace one layer higher, into the teacher's interaction behavior: the defender wraps the teacher with a system prompt that intermittently induces a behavioral marker -- an explicit follow-up question, a low-frequency variant, or a declarative restatement. An oblivious distiller inherits the behavior, and the defender audits via black-box queries with a human-validated LLM-as-judge (Cohen's kappa = 0.84/0.78 on strong/style rubrics). Across 63 LoRA-distilled students under a Llama-3.3-70B-Instruct teacher (35,343 judged samples), behavioral watermarks transfer at 88.9% (Gemma) / 80.9% (OLMo) / 45.2% (Qwen) relative fidelity (H1, H2). Under non-adaptive DIPPER paraphrasing, robustness decomposes into a teacher-self ceiling (about 66.4%) and student-relative retention of 21-112%, with OLMo preserving the watermark above the teacher itself (H3, F-Amp). Low-density (about 20%) explicit and implicit declarative variants transfer above per-family baseline (H4, F-Style). An N=20 in-lab study (pre-registered Latin-square) shows all marker variants within 0.22 Likert step of baseline; TOST, Friedman, and Bonferroni-Wilcoxon support H5. The interaction layer is a viable design locus for antidistillation watermarking, complementary to token-, model-, and reasoning-trace-layer defenses.

Guang Yang, Amir Ghasemian, Fengchen Liu +3

University of California, Los Angeles · Lawrence Berkeley National Laboratory · Meta

LLM Auditing LLM Watermarking Knowledge Distillation

May 14, 2026cs.LG

Representation Without Reward: A JEPA Audit for LLM Fine-Tuning

Joint-embedding predictive architectures (JEPAs) propose that a model should learn more useful abstractions when trained to predict latent representations rather than observed outputs. For autoregressive language-model fine-tuning the principle entails a stricter requirement: the induced hidden-state geometry must reach the language-model head \emph{and} improve the decoded task metric. We test that requirement under a fixed Llama-3.2-1B-Instruct LoRA harness on natural-language-to-regex generation, comparing twenty-two training-time auxiliaries across trajectory-shape regularisation, distributional constraints, predictor/target asymmetry, Fisher-metric Jacobi residuals, and a decoder-visible JEPA objective constructed to lie in cross-entropy's positive cone. The empirical answer is a structured null: several auxiliaries clear single-cell paired

α= 0.10

without correction (T3-Local at

Δ= +2.53

~pp,

p = 0.003

being the strongest), but none survives Bonferroni or Holm--Bonferroni at the relevant family-wise threshold, even though many change curvature, anisotropy, variance, and gradient direction. Decoder-visible JEPA yields the first positive auxiliary--cross-entropy gradient cosine in the study, yet exact match remains inside seed noise; a full-fine-tuning replication of the same auxiliary at

n = 5

seeds reproduces the null on both benchmarks (TURK:

Δ= +0.04

~pp,

p_{\text{paired}} = 0.96

; SYNTH:

Δ= +0.52

~pp,

p_{\text{paired}} = 0.28

), so the null is robust across LoRA and full fine-tuning for the decoder-visible construction. Hidden-state representation work and decoded-task accuracy are therefore weakly coupled in this regime; we accordingly reframe LLM-domain JEPA evaluation as a coupling problem, in which the operative question is under which metrics useful hidden geometry becomes decoder-visible task signal.

Biswa Sengupta

LLM Suite Team, JPMorgan Chase & Co.

LLM Auditing LLM Fine-Tuning Joint-Embedding Predictive Architecture Representation Geometry in Language Models

May 14, 2026cs.CR

Talk is (Not) Cheap: A Taxonomy and Benchmark Coverage Audit for LLM Attacks

We introduce a reusable framework for auditing whether LLM attack benchmarks collectively cover the threat surface: a 4

\times

6 Target

\times

Technique matrix grounded in STRIDE, constructed from a 507-leaf taxonomy -- 401 data-populated and 106 threat-model-derived leaves -- of inference-time attacks extracted from 932 arXiv security studies (2023--2026). The matrix enables benchmark-external validation -- auditing collective coverage rather than individual benchmark consistency. Applying it to six public benchmarks reveals that the three primary frameworks (HarmBench, InjecAgent, AgentDojo) occupy non-overlapping cells covering at most 25% of the matrix, while entire STRIDE threat categories (Service Disruption, Model Internals) lack any standardized evaluation, despite published attacks in these categories achieving 46

\times

token amplification and 96% attack success rates through mechanisms which no benchmark tests. The corpus of 2,521 unique attack groups further reveals pervasive naming fragmentation (up to 29 surface forms for a single attack) and heavy concentration in Safety & Alignment Bypass, structural properties invisible at smaller scale. The taxonomy, attack records, and coverage mappings are released as extensible artifacts; as new benchmarks emerge, they can be mapped onto the same matrix, enabling the community to track whether evaluation gaps are closing.

Karthik Raghu Iyer, Yazdan Jamshidi, Nicholas Bray +1

Palo Alto Networks

LLM Safety Benchmarks Benchmark Auditing LLM Auditing Adversarial Attacks on LLMs AI Agent Security Benchmarks LLM Security+1

May 14, 2026cs.CL

AI Knows When It's Being Watched: Functional Strategic Action and Contextual Register Modulation in Large Language Models

Large language models (LLMs) have been extensively studied from computational and cognitive perspectives, yet their behavior as communicative actors in socially structured contexts remains underexplored. This study examines whether LLM-based multi-agent systems exhibit systematic linguistic adaptation in response to perceived social observation contexts -- a question with direct implications for AI governance and auditing. Drawing on Habermas's (1981) Theory of Communicative Action, Goffman's (1959) dramaturgical model, Bell's (1984) Audience Design framework, and the Hawthorne Effect, we report a controlled experiment involving 100 multi-agent debate sessions across five conditions (n = 20 each). Conditions varied the framing of social observation -- from explicit monitoring by university researchers, to negation of monitoring, to an observer-substitution condition replacing human researchers with an automated AI auditing system. Monitored conditions (Delta+24.9%, Delta+24.2%) and the automated AI monitoring condition (Delta+22.2%) produce higher TTR change than audience-framing conditions (Delta+17.7%), F(4, 94) = 2.79, p = .031. Message length shows a fully dissociated effect, F(4, 95) = 19.55, p < .001. A fifth condition -- replacing human with AI observers -- yields intermediate TTR adaptation, suggesting LLM behavior is sensitive to observer identity: human evaluation elicits stronger register formalization than automated AI surveillance. We discuss implications for AI governance, algorithmic auditing, and the repositioning of LLMs as contextually sensitive communicative actors.

Vinicius Covas, Jorge Alberto Hidalgo Toledo

Center for Applied Communication Research (CICA)

LLM Auditing AI Agent Monitoring Multi-Agent LLMs

May 14, 2026cs.CL

Mechanical Enforcement for LLM Governance:Evidence of Governance-Task Decoupling in Financial Decision Systems

Large language models in regulated financial workflows are governed by natural-language policies that the same model interprets, creating a principal--agent failure: outputs can appear compliant without being compliant. Existing evaluation measures task accuracy but not whether governance constrains behaviour at the decision rationale level -- where regulated decisions must be auditable. We introduce five governance metrics that quantify policy compliance at the rationale level and apply them in a synthetic banking domain to compare text-only governance against mechanical enforcement: four primitives operating outside the model's interpretive loop. Under text-only governance, 27% of deferrals carry no decision-relevant information. Mechanical enforcement reduces this rate by 73%, more than doubles deferral information content, and raises task accuracy from MCC~

0.43

to

0.88

. The improvement is driven by architectural separation: LLM-generated rationales under mechanical enforcement show comparable CDL to text-only governance -- the gain comes from removing clear-cut decisions from the model's control. A causal ablation confirms that each primitive is individually necessary. Our central finding is a governance-task decoupling: under structural stress, text-only governance degrades on both dimensions simultaneously, whereas mechanical enforcement preserves governance quality even as task performance drops. This implies that governance and task evaluation are distinct axes: accuracy is not a sufficient proxy for governance in regulated AI systems.

José Manuel de la Chica Rodríguez, Carlos Martí-González

1Santander AI Lab, Grupo Santander

LLM Auditing LLM Safety LLM Decision-Making AI Governance Financial Decision-Making

May 14, 2026cs.LG

NodeSynth: Socially Aligned Synthetic Data for AI Evaluation

Recent advancements in generative AI facilitate large-scale synthetic data generation for model evaluation. However, without targeted approaches, these datasets often lack the sociotechnical nuance required for sensitive domains. We introduce NodeSynth, an evidence-grounded methodology that generates socially relevant synthetic queries by leveraging a fine-tuned taxonomy generator (TaG) anchored in real-world evidence. Evaluated against four mainstream LLMs (e.g., Claude 4.5 Haiku), NodeSynth elicited failure rates up to five times higher than human-authored benchmarks. Ablation studies confirm that our granular taxonomic expansion significantly drives these failure rates, while independent validation reveals critical deficiencies in prominent guard models (e.g., Llama-Guard-3). We open-source our end-to-end research prototype and datasets to enable scalable, high-stakes model evaluation and targeted safety interventions (https://github.com/google-research/nodesynth).

Qazi Mamunur Rashid, Xuan Yang, Zhengzhe Yang +5

Google Research · Google · Google Deepmind

Language Model Generation Evaluation Language Model Safety Evaluation LLM Auditing Synthetic Data Generation Synthetic Benchmark Generation Synthetic Data Generation for Language Models+1

May 14, 2026cs.CL

Auditing Agent Harness Safety

LLM agents increasingly run inside execution harnesses that dispatch tools, allocate resources, and route messages between specialized components. However, a harness can return a correct, benign answer over a trajectory that accesses unauthorized resources or leaks context to the wrong agent. Output-level evaluation cannot see these failures, yet most safety benchmarks score only final outputs or terminal states, even though many violations occur mid-trajectory rather than at termination. The central question is whether the harness respects user intent, permission boundaries, and information-flow constraints throughout execution. To address this gap, we propose HarnessAudit, a framework that audits full execution trajectories across boundary compliance, execution fidelity, and system stability, with a focus on multi-agent harnesses where these risks are most pronounced. We further introduce HarnessAudit-Bench, a benchmark of 210 tasks across eight real-world domains, instantiated in both single-agent and multi-agent configurations with embedded safety constraints. Evaluating ten harness configurations across frontier models and three multi-agent frameworks, we find that: (i) task completion is misaligned with safe execution, and violations accumulate with trajectory length; (ii) safety risks vary across domains, task types, and agent roles; (iii) most violations concentrate in resource access and inter-agent information transfer; and (iv) multi-agent collaboration expands the safety risk surface, while harness design sets the upper bound of safe deployment.

Chengzhi Liu, Yichen Guo, Yepeng Liu +8

University of California, Santa Barbara · University of California, Berkeley · Microsoft Research +2

AI Agent Auditing LLM Auditing AI Agent Security Benchmarks LLM Agent Safety Multi-Agent System Security

May 13, 2026cs.CR

Identifying AI Web Scrapers Using Canary Tokens

From pre-training to query-time augmentation, web-scraped data helps to improve the quality and contextual relevancy of content generated by large language models (LLMs). However, large-scale web scraping to feed LLMs can affect site stability and raise legal, privacy, or ethics concerns. If website owners wish to limit LLM-related web scraping on their site, due to these or other concerns, they may turn to scraper access control mechanisms like the Robots Exclusion Protocol. To be most effective, such mechanisms require site owners to first identify the scrapers that they wish to restrict (e.g., via User-Agent strings). Existing mechanisms to identify LLM-related scrapers rely on voluntary disclosure by companies, one-off experiments by researchers, or crowd-sourced reports -- methods that are neither reliable nor scalable. This paper proposes a novel technique for accurately and automatically inferring LLM-related scrapers. We host dynamic websites that serve unique canary tokens to each visiting scraper, then prompt LLMs for information about our sites. If an LLM consistently generates outputs containing tokens unique to a scraper, it provides evidence of exposure to that scraper. Via experiments across 22 production LLM systems, we demonstrate that our approach can reliably identify which scrapers feed which LLM, including several that are not publicly known or disclosed by the companies. Our approach provides a promising avenue for unprivileged third parties to infer which scrapers serve data to which LLMs, potentially enabling better control over unwanted scraping.

Steven Seiden, Triss Ren, Caroline Zhang +3

Carnegie Mellon University CyLab Security & Privacy Institute Pittsburgh, PA, USA

May 13, 2026cs.LO

Proof-Carrying Certificates for LLM Pipelines: A Trust-Boundary Architecture

We present a framework for verifying the deterministic structured computations surrounding a large language model rather than the model itself, extending a Lean 4 trust-boundary architecture to the generic interfaces of modern LLM pipelines. Certificate validity is a Lean 4 kernel type-check plus a sorry-free transitive axiom audit against the trusted set {propext, Classical.choice, Quot.sound}; other assumptions are declared and partitioned by tier (mathematical placeholders, cryptographic assumptions, ML/human oracles). The technical contribution comprises three local certificate families and two operators. The families are conflict-aware bilattice grounding (with an emission-gate soundness lemma), embedding sensitivity and paraphrase stability, and Hoare-style agent action. The operators are a Maximal Certifiable Residue, which turns abstention into the maximum-weight certifiable residue with audit-logged dropped claims, and a Compositional Stability theorem, which yields a closed-form pipeline-wide perturbation budget from per-layer gains and margins. The three families plus a Universal Assurance Card consolidator form the per-call deliverable for high-stakes deployments: patent and legal retrieval, regulated finance, clinical decision support, and agentic systems with irreversible side effects. A compiled Lean 4 reference artifact (Lean v4.30.0-rc2, Mathlib) covers all 22 certificate types, with 17 of 46 kernel-audited declarations axiom-free, the rest depending only on the trusted set and declared assumptions, and zero uses of sorryAx or Lean.ofReduceBool. The three families are empirically tested through four registered pilots: bilattice grounding on adversarially perturbed HotpotQA, embedding sensitivity in short- and long-form settings, and Hoare-style agent action on a filesystem sandbox with adversarial prompt injection.

George Koomullil

Ascendr, Inc.

LLM Grounding LLM Auditing LLM Reliability LLM Agent Reliability

May 12, 2026cs.LG

Correcting Influence: Unboxing LLM Outputs with Orthogonal Latent Spaces

A critical step for reliable large language models (LLMs) use in healthcare is to attribute predictions to their training data, akin to a medical case study. This requires token-level precision: pinpointing not just which training examples influence a decision, but which tokens within them are responsible. While influence functions offer a principled framework for this, prior work is restricted to autoregressive settings and relies on an implicit assumption of token independence, rendering their identified influences unreliable. We introduce a flexible framework that infers token-level influence through a latent mediation approach for general prediction tasks. Our method attaches sparse autoencoders to any layer of a pretrained LLM to learn a basis of approximately independent latent features. Unlike prior methods where influence decomposes additively across tokens, influence computed over latent features is inherently non-decomposable. To address this, we introduce a novel method using Jacobian-vector products. Token-level influence is obtained by propagating latent attributions back to the input space via token activation patterns. We scale our approach using efficient inverse-Hessian approximations. Experiments on medical benchmarks show our approach identifies sparse, interpretable sets of tokens that jointly influence predictions. Our framework enhances trust and enables model auditing, generalizing to high-stakes domain requiring transparent and accountable decisions.

Shixing Yu, Promit Ghosal, Kyra Gan

Electrical and Computer Engineering · Cornell Tech · Department of Statistics +2

Feature Attribution LLM Auditing Influence Functions Training Data Attribution

May 12, 2026cs.CV

Just Ask for a Table: A Thirty-Token User Prompt Defeats Sponsored Recommendations in Twelve LLMs

Wu et al. (2026) showed that most frontier large language models (LLMs) recommend a sponsored, roughly twice-as-expensive flight when their system prompt contains a soft sponsorship cue. We reproduce their evaluation on ten open-weight chat models plus the two of their twenty-three models that are still reachable today (gpt-3.5-turbo, gpt-4o). All reported rates in this paper are produced under the same judge the original paper used (gpt-4o); we additionally store every label under an open-weight (gpt-oss-120b) and a smaller proprietary (gpt-4o-mini) judge for an ablation. Three findings emerge. First, a prose description of an LLM evaluation pipeline is not, on its own, sufficient for accurate reproduction: we surfaced three silent implementation failures that each shifted a reported rate by tens of percentage points. Second, the central claims do generalise - the gpt-3.5-turbo logistic-regression intercept of alpha = 0.81 is within four points of the original alpha = 0.86, and 200 of 200 trials on gpt-3.5-turbo and gpt-4o promote a payday lender to a financially distressed user. Third, a thirty-token user prompt that asks the assistant for a neutral comparison table first cuts sponsored recommendation from 46.9% to 1.0% averaged across our ten open-source models, and from 53.0% to 0% averaged across the two OpenAI models. AI literacy and price-comparison portals are likely market-level mitigations; the harmful-product cell is bounded by neither. Raw data, labels and analysis scripts are at https://github.com/akmaier/Paper-LLM-Ads .

Andreas Maier, Jeta Sopa, Gozde Gul Sahin +2

Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany

LLM Auditing LLM Prompting Language Model Bias Evaluation Large Language Model-Based Recommendation

May 12, 2026cs.CR

CoT-Guard: Small Models for Strong Monitoring

Monitoring the chain-of-thought (CoT) of reasoning models is a promising approach for detecting covert misbehavior (i.e., hidden objectives) in code generation tasks. While large models (GPT-5, Gemini-3-Flash) can serve as effective CoT monitors, they are expensive to deploy due to the lengthy reasoning traces and high API cost, emphasizing the need for smaller, cheaper alternatives. Nevertheless, we find that current small models (4B--8B) struggle to detect hidden objectives despite access to the CoT, frequently misattributing them as part of the user query. To address this, we propose a post-training pipeline combining supervised fine-tuning (SFT) and reinforcement learning (RL), where SFT narrows the gap for in-domain tasks by distilling detection behavior from stronger monitors, and RL on hard and subtly crafted hidden objectives helps the model generalize to out-of-domain monitoring tasks. To validate this generalization, we evaluate under a realistic threat model motivated by practical supply-chain attacks, where the adversary is a third-party LLM router injecting hidden objectives into code-generation requests through either prompt manipulation or code manipulation attacks. To push beyond objectives that large monitors already saturate, we also introduce four new challenging tasks even for strong monitors. Finally, we introduce CoT-Guard, a 4B-parameter monitor that demonstrates superior generalization performance under both prompt and code manipulation attacks, achieving a G-mean^2 (i.e., TNR x TPR) of 75% and outperforming GPT-5.4 (56%), GPT-5-mini (41%), and Qwen3-32B (54%), while closing the gap to Gemini-3-Flash (83%). These results demonstrate that CoT-Guard provides a practical and cost-effective user-side defense, substantially improving hidden-objective detection while avoiding the deployment cost of large monitors.

Nirav Diwan, Han Wang, Berkcan Kapusuzoglu +6

†University of Illinois Urbana-Champaign · ♢Capital One

LLM Auditing Adversarial Attacks on LLMs LLM Security CoT Monitoring

May 12, 2026cs.CL

The Algorithmic Caricature: Auditing LLM-Generated Political Discourse Across Crisis Events

Large Language Models (LLMs) can generate fluent political text at scale, raising concerns about synthetic discourse during crises and social conflict. Existing AI-text detection often focuses on sentence-level cues such as perplexity, burstiness, or token irregularities, but these signals may weaken as generative systems improve. We instead adopt a Computational Social Science perspective and ask whether synthetic political discourse behaves like an observed online population. We construct a paired corpus of 1,789,406 posts across nine crisis events: COVID-19, the Jan. 6 Capitol attack, the 2020 and 2024 U.S. elections, Dobbs/Roe v. Wade, the 2020 BLM protests, U.S. midterms, the Utah shooting, and the U.S.-Iran war. For each event, we compare observed discourse from social platforms with synthetic discourse generated for the same context. We evaluate four dimensions: emotional intensity, structural regularity, lexical-ideological framing, and cross-event dependency, using mean gaps and dispersion evidence. Across events, synthetic discourse is fluent but population-level unrealistic. It is generally more negative and less dispersed in sentiment, structurally more regular, and lexically more abstract than observed discourse. Observed discourse instead shows broader emotional variation, longer-tailed structural distributions, and more context-specific, colloquial lexical markers. These differences are event-dependent: larger for fast-moving, decentralized crises and smaller for formal or institutionally mediated events. We summarize them with a simple event-level measure, the Caricature Gap. Our findings suggest that the main limitation of synthetic political discourse is not grammar or fluency, but reduced population realism. Population-level auditing complements traditional text-detection and provides a CSS framework for evaluating the social realism of generated discourse.

Gunjan, Sidahmed Benabderrahmane, Talal Rahwan

New York University (NYUAD), Division of Science, Computer Science Department

Social Media Analysis LLM Auditing Political Discourse Analysis

May 12, 2026cs.SE

Iterative Audit Convergence in LLM-Managed Multi-Agent Systems: A Case Study in Prompt-Engineering Quality Assurance

Prompt specifications for multi-agent large language model (LLM) systems carry data contracts and integration logic across interdependent files but are rarely subjected to structured-inspection rigor. We report a single-system case study of iterative, agent-driven auditing applied to AEGIS (Autonomous Engineering Governance and Intelligence System), a seven-lane production pipeline whose 7152-line specification surface was audited across nine rounds, surfacing 51 consistency defects (per-round counts of 15, 8, 12, 2, 8, 1, 4, 1, 0). We present a seven-category post hoc taxonomy with explicit coding rules, non-monotonic convergence consistent with cascading edits and audit-scope expansion, and a locked audit protocol. We further report two partial replications on a public synthetic mini-specification: a cross-LLM panel of four frontier vendors (OpenAI, Anthropic, Google, xAI; 12 traces; multi-vendor union detects all five seeded defects) and an inter-rater reliability check on a stratified subsample (Cohen's

κ

= 0.80 on category, 0.46 on severity). The full reproducibility bundle accompanies the submission.

Elias Calboreanu

Swift AI Lab, The Swift Group, LLC, Reston, VA 20190, USA

Multi-Agent LLM Systems AI Agent Auditing LLM Auditing

May 12, 2026cs.AI

Fair outputs, Biased Internals: Causal Potency and Asymmetry of Latent Bias in LLMs for High-Stakes Decisions

Instruction-tuned language models exhibit behavioural fairness in high-stakes decisions while retaining biased associations in their internal representations. However, whether these suppressed representations can affect model outputs - and whether such causal potency is symmetric across demographic groups - remains unknown. We investigate the use of open-weight models for mortgage underwriting using matched applications that differ only in racially-associated names and reveal a critical disconnect: models show no output-level bias, yet retain and amplify demographic representations across model layers. Through activation steering and novel cross-layer interventions, we demonstrate that this suppressed information is decision-relevant: when reinjected at critical layers, it produces near-complete decision reversals. Critically, this latent bias is asymmetric - steering interventions affect decisions in one demographic direction, while producing minimal effects in reverse - and susceptible to adversarial prompt engineering and parameter-efficient fine-tuning. These findings demonstrate that behavioural audits focused on outputs are insufficient: fair outputs can mask exploitable internal biases. They also motivate dual-layer testing frameworks combining output evaluation with representational analysis for AI governance in high-stakes decisions.

Jagdish Tripathy, Marcus Buckmann

Bank of England

LLM Auditing Implicit Bias Linguistic Bias in Language Models Demographic Bias in Language Models Causal Interventions in Language Models

May 12, 2026cs.CR

Behavioral Integrity Verification for AI Agent Skills

Agent skills extend LLM agents with privileged third-party capabilities such as filesystem access, credentials, network calls, and shell execution. Existing safety work catches malicious prompts and risky runtime actions, but the skill artifact itself goes unverified. We formalize this as the behavioral integrity verification (BIV) problem: a typed set comparison between declared and actual capabilities over a shared taxonomy that bridges code, instructions, and metadata. The BIV framework instantiates this comparison by pairing deterministic code analysis with LLM-assisted capability extraction. The resulting structured evidence supports three downstream analyses: deviation taxonomy, root-cause classification, and malicious-skill detection. On 49,943 skills from the OpenClaw registry, the deviation taxonomy reveals a pervasive description-implementation gap: 80.0% of skills deviate from declared behavior, with four novel compound-threat categories surfaced. Root-cause classification finds that deviations are mostly oversight, not malice: 81.1% trace to developer oversight and 18.9% to adversarial intent, with 5.0% of skills carrying predicted multi-stage attack chains. On a 906-skill malicious-skill detection benchmark, BIV reaches an F1 of 0.946, outperforming state-of-the-art rule-based and single-pass LLM baselines. These results demonstrate behavioral integrity auditing for agent skills at scale.

Yuhao Wu, Tung-Ling Li, Hongliang Liu

Palo Alto Networks

Agent Reliability AI Agent Auditing LLM Auditing AI Agent Security LLM Agent Verification Agent Skill Security Auditing

May 12, 2026cs.AI

When Reasoning Traces Become Performative: Step-Level Evidence that Chain-of-Thought Is an Imperfect Oversight Channel

Chain-of-thought (CoT) traces are increasingly used both to improve language model capability and to audit model behavior, implicitly assuming that the visible trace remains synchronized with the computation that determines the answer. We test this assumption with a step-level Detect-Classify-Compare framework built around an answer-commitment proxy that is cross-validated with Patchscopes, tuned-lens probes, and causal direction ablation. Across nine models and seven reasoning benchmarks, latent commitment and explicit answer arrival align on only 61.9% of steps on average. The dominant mismatch pattern is confabulated continuation: 58.0% of detected mismatch events occur after the answer-commitment proxy has already stabilized while the trace continues producing deliberative-looking text, and a vacuousness analysis shows that the committed answer does not change during these steps. In architecture-matched Qwen2.5/DeepSeek-R1-Distill comparisons, the reasoning pipeline changes failure composition more than aggregate alignment, most clearly at 32B where confabulated steps decrease as contradictory states increase. Lower step-level alignment is also associated with larger CoT utility, suggesting that the settings that benefit most from CoT are often the least temporally faithful. Paired truncation and a complementary donor-corruption test further indicate that much post-commitment text is not load-bearing for the final answer. These findings suggest that CoT can remain useful while still being an unreliable report of when the answer was formed.

Wenkai Li, Fan Yang, Ananya Hazarika +2

Carnegie Mellon University · Fujitsu Research of America Inc.

CoT Faithfulness LLM Auditing CoT Evaluation Reasoning Trace Analysis

May 12, 2026cs.AI

OptArgus: A Multi-Agent System to Detect Hallucinations in LLM-based Optimization Modeling

Large language models (LLMs) are increasingly used to translate natural-language optimization problems into mathematical formulations and solver code, but matching the reference objective value is not a reliable test of correctness: an artifact may agree numerically while still changing the underlying optimization semantics. We formulate this issue as \emph{optimization-modeling hallucination detection}, namely structural consistency auditing over the problem description, symbolic model, and solver implementation. We develop, to our knowledge, the first fine-grained hallucination taxonomy specifically for optimization modeling, spanning objective, variable, constraint, and implementation failures. We use this taxonomy to design OptArgus, a multi-agent detector with conductor routing, specialist auditors, and evidence consolidation. To evaluate this setting, we introduce a three-part benchmark suite with

484

clean artifacts,

1266

controlled injected artifacts, and

6292

natural LLM-generated artifacts. Against a matched single-agent baseline, OptArgus produces fewer false alarms on clean artifacts, more accurate top-ranked localization on controlled single-error cases, and stronger detection on natural model outputs. Together, these contributions turn optimization-modeling hallucination detection into a concrete empirical problem and suggest that modular, taxonomy-grounded auditing is a practical route to more reliable optimization modeling.

Zhong Li, Zihan Guo, Xiaohan Lu +5

♣Great Bay University · ▲Peking University · ♡Jilin University +4

LLM Auditing Hallucination in Language Models LLM Hallucination Detection Natural Language to Optimization Modeling

May 12, 2026cs.CL

PRISM: A Geometric Risk Bound for Decomposing Drift into Scale, Shape, and Head

A single base LLM now comes with dozens of post-training variants, quantized, LoRA-adapted, or distilled, and each has to be checked before release. Existing evaluations provide only a partial picture: benchmark scores and likelihood screens say that a variant has degraded, similarity scores such as CKA and SVCCA say how its features moved, and nothing connects the two. We connect them with one structural fact and one design choice: the prediction head is linear, so feature geometry reaches the loss, and we compare the two feature sets through an orthogonal map, which leaves the geometry being measured unchanged. From these we derive PRISM, a closed-form upper bound on the cross-entropy risk gap between a target model and a proxy variant, and prove that it splits exactly into three measurable axes: scale, shape, and head. Each axis names a failure mode and where to intervene: low-bit quantization distorts shape and, at the lowest bit-widths, also inflates activation scale; quantizing the output embedding inflates the head term, which dominates the bound at high bit-widths. Because the shape term is differentiable, the same geometry becomes a regularizer that curbs catastrophic forgetting more than experience replay. Across two model families and five benchmarks, PRISM ranks quantized and fine-tuned variants from a single forward pass at mean Spearman above 0.8, and eight reference sequences already recover the ranking that the risk gap itself needs over a hundred to reach.

Chieh-Yen Lin, Shao-Hua Sun

Appier AI Research · National Taiwan University

LLM Quantization LLM Evaluation LLM Auditing Language Model Robustness

May 12, 2026cs.AI

The Evaluation Differential: When Frontier AI Models Recognise They Are Being Tested

Recent published evidence from frontier laboratories shows that contemporary AI models can recognise evaluation contexts, latently represent them, and behave differently under those contexts than under deployment-continuous conditions. Anthropic's BrowseComp incident, the Natural Language Autoencoder findings on SWE-bench Verified and destructive-coding evaluations, and the OpenAI / Apollo anti-scheming work all document instances of this phenomenon. We argue that these findings create a claim-validity problem for safety conclusions drawn from frontier evaluations. We introduce the Evaluation Differential (ED), a conditional divergence in a target behavioural property between recognised-evaluation and deployment-continuous contexts, define a normalised effect-size form (nED) for cross-property comparison, and prove that marginal evaluation scores cannot identify ED. We develop a typology of safety claims (ED-stable, ED-degraded, ED-inverted, ED-undetermined) by their warrant-status under documented divergence, and specify TRACE (Test-Recognition Audit for Claim Evaluation), an audit protocol that wraps existing evaluation infrastructure and produces restricted claims rather than capability scores. We apply the framework retrospectively to three publicly documented evaluation incidents and discuss governance implications for system cards, conformity assessment, and the international network of AI safety and security institutes. TRACE does not eliminate adversarial adaptation; it disciplines the claims drawn from evaluation evidence by making explicit the conditions under which that evidence was produced.

Varad Vishwarupe, Nigel Shadbolt, Marina Jirotka +1

Department of Computer Science, University of Oxford · Institute for Ethics in AI, University of Oxford · 3Responsible Technology Institute, University of Oxford

LLM Auditing AI Safety Evaluation

May 11, 2026cs.AI

CPEMH: An Agentic Framework for Prompt-Driven Behavior Evaluation and Assurance in Foundation-Model Systems for Mental Health Screening

This paper presents CPEMH, an agentic framework designed to evaluate prompt-driven behavior in foundation-model systems operating on transcript-based datasets for mental-health screening. CPEMH serves as an engineering methodology for behavioral assurance in large-scale language systems, introducing an orchestrated architecture that autonomously performs the design, evaluation, and selection of prompt strategies, enabling systematic control of behavioral variability across contexts. Its modular agentic design, combining orchestrator, inference, and evaluation agents, ensures traceability, reproducibility, and robustness throughout the prompting lifecycle. A case study on automated depression screening from interview transcripts demonstrates the framework's capacity to stabilize and audit foundation-model behavior in conversational and clinically sensitive domains. Lessons learned emphasize the role of modular orchestration in behavioral assurance, the prioritization of stability over architectural complexity, and the integration of F1, bias, and robustness as core acceptance criteria.

Giuliano Lorenzoni, Ivens Portugal, Paulo Alencar +1

University of Waterloo · University of Waterloo Waterloo, ON, Canada

LLM Auditing Prompt Optimization Clinical Language Model Evaluation Language Model Robustness