LLM Auditing

LLM: Large Language Model

Overview

Large language model auditing examines how models are trained and how they behave in use, looking for problems such as bias, privacy leakage, hallucinations, unsafe actions, and unreliable reasoning. Common approaches include controlled benchmarks and counterfactual tests, multi-turn red-teaming, membership-inference and provenance analysis, human review, and probes of model activations or weights. Audits increasingly examine whole systems—including prompts, tools, interfaces, and model updates—rather than judging only a model’s final answer. A central challenge is that results can shift with prompts, access methods, versions, and evaluators, while behavioral scores or internal signals alone may not establish why a failure occurred or whether it will recur in deployment.

Latest papers

723 papers79 papers in the last four weeks, down 15% on the four weeks before. 0.6% of all new papers.

Mar 17, 2026

Detecting Data Poisoning in Code Generation LLMs via Black-Box, Vulnerability-Oriented Scanning

Code generation large language models (LLMs) are increasingly integrated into modern software development workflows. Recent work has shown that these models are vulnerable to backdoor and poisoning attacks that induce the generation of insecure code, yet effective defenses remain limited. Existing scanning approaches rely on token-level generation consistency to invert attack targets, which is ineffective for source code where identical semantics can appear in diverse syntactic forms. We present CodeScan, the first black-box, vulnerability-specific scanning framework for auditing code generation LLMs, assuming that the defender specifies the target vulnerability classes and provides corresponding task-relevant prompts. CodeScan identifies attack targets by analyzing structural similarities across multiple generations conditioned on different clean prompts. It combines iterative divergence analysis with abstract syntax tree (AST)-based normalization to abstract away surface-level variation and unify semantically equivalent code, isolating structures that recur consistently across generations. CodeScan then applies LLM-based vulnerability analysis to determine whether the extracted structures contain security vulnerabilities and flags the model as compromised when such a structure is found. We evaluate CodeScan against four representative attacks under both backdoor and poisoning settings across three real-world vulnerability classes. Experiments on 117 models spanning three architectures and multiple model sizes demonstrate 97%+ detection accuracy with substantially lower false positives than prior methods.
Mar 17, 2026

Are Large Language Models Truly Smarter Than Humans?

Public leaderboards increasingly suggest that large language models (LLMs) surpass human experts on benchmarks spanning academic knowledge, law, and programming. Yet most benchmarks are fully public, their questions widely mirrored across the internet, creating systematic risk that models were trained on the very data used to evaluate them. This paper presents three complementary experiments forming a rigorous multi-method contamination audit of six frontier LLMs: GPT-4o, GPT-4o-mini, DeepSeek-R1, DeepSeek-V3, Llama-3.3-70B, and Qwen3-235B. Experiment 1 applies a lexical contamination detection pipeline to 513 MMLU questions across all 57 subjects, finding an overall contamination rate of 13.8% (18.1% in STEM, up to 66.7% in Philosophy) and estimated performance gains of +0.030 to +0.054 accuracy points by category. Experiment 2 applies a paraphrase and indirect-reference diagnostic to 100 MMLU questions, finding accuracy drops by an average of 7.0 percentage points under indirect reference, rising to 19.8 pp in both Law and Ethics. Experiment 3 applies TS-Guessing behavioral probes to all 513 questions and all six models, finding that 72.5% trigger memorization signals far above chance, with DeepSeek-R1 displaying a distributed memorization signature (76.6% partial reconstruction, 0% verbatim recall) that explains its anomalous Experiment 2 profile. All three experiments converge on the same contamination ranking: STEM > Professional > Social Sciences > Humanities.
Mar 16, 2026

Invisible Influences: Investigating Implicit Intersectional Biases through Persona Engineering in Large Language Models

Large Language Models (LLMs) excel at human-like language generation but often embed and amplify implicit, intersectional biases, especially under persona-driven contexts. Existing bias audits rely on static, embedding-based tests (CEAT, I-WEAT, I-SEAT) that quantify absolute association strengths. We show that they have limitations in capturing dynamic shifts when models adopt social roles. We address this gap by introducing the Bias Amplification Differential and Explainability Score (BADx): a novel, scalable metric that measures persona-induced bias amplification and integrates local explainability insights. BADx comprises three components - differential bias scores (BAD, based on CEAT, I-WEAT, I-SEAT),Persona Sensitivity Index (PSI), and Volatility (Standard Deviation), augmented by LIME-based analysis for emphasizing explainability. This study is divided and performed as two different tasks. Task 1 establishes static bias baselines, and Task 2 applies six persona frames (marginalized and structurally advantaged) to measure BADx, PSI, and volatility. This is studied across five state-of-the-art LLMs (GPT-4o, DeepSeek-R1, LLaMA-4, Claude 4.0 Sonnet and Gemma-3n E4B). Results show persona context significantly modulates bias. GPT-4o exhibits high sensitivity and volatility; DeepSeek-R1 suppresses bias but with erratic volatility; LLaMA-4 maintains low volatility and a stable bias profile with limited amplification; Claude 4.0 Sonnet achieves balanced modulation; and Gemma-3n E4B attains the lowest volatility with moderate amplification. BADx performs better than static methods by revealing context-sensitive biases overlooked in static methods. Our unified method offers a systematic way to detect dynamic implicit intersectional bias in five popular LLMs.
Mar 16, 2026

InterPol: De-anonymizing LM Arena via Interpolated Preference Learning

Strict anonymity of model responses is a key for the reliability of voting-based leaderboards, such as LM Arena. While prior studies have attempted to compromise this assumption using simple statistical features like TF-IDF or bag-ofwords, these methods often lack the discriminative power to distinguish between stylistically similar or within-family models. To overcome these limitations and expose the severity of vulnerability, we introduce INTERPOL, a model-driven identification framework that learns to distinguish target models from others using interpolated preference data. Specifically, INTERPOL captures deep stylistic patterns that superficial statistical features miss by synthesizing hard negative samples through model interpolation and employing an adaptive curriculum learning strategy. Extensive experiments demonstrate that INTERPOL significantly outperforms existing baselines in identification accuracy. Furthermore, we quantify the real-world threat of our findings through ranking manipulation simulations on Arena battle data.
Mar 15, 2026

Seamless Deception: Larger Language Models Are Better Knowledge Concealers

Language Models (LMs) may acquire harmful knowledge, and yet feign ignorance of these topics when under audit. Inspired by the recent discovery of deception-related behaviour patterns in LMs, we aim to train classifiers that detect when a LM is actively concealing knowledge. Initial findings on smaller models show that classifiers can detect concealment more reliably than human evaluators, with gradient-based concealment proving easier to identify than prompt-based methods. However, contrary to prior work, we find that the classifiers do not reliably generalize to unseen model architectures and topics of hidden knowledge. Most concerningly, the identifiable traces associated with concealment become fainter as the models increase in scale, with the classifiers achieving no better than random performance on any model exceeding 70 billion parameters. Our results expose a key limitation in black-box-only auditing of LMs and highlight the need to develop robust methods to detect models that are actively hiding the knowledge they contain.
Mar 15, 2026

AEX: Non-Intrusive Multi-Hop Attestation and Provenance for LLM APIs

Hosted large language models are increasingly accessed through remote APIs, but the API boundary still offers little direct evidence that a returned output actually corresponds to the client-visible request. Recent audits of shadow APIs show that unofficial or intermediary endpoints can diverge from claimed behavior, while existing approaches such as fingerprinting, model-equality testing, verifiable inference, and TEE attestation either remain inferential or answer different questions. We propose AEX, a non-intrusive attestation extension for existing JSON-based LLM APIs. AEX preserves request, response, tool-calling, streaming, and error semantics, and instead adds a signed top-level attestation object that binds a client-visible request projection to either a complete response object or a committed streaming output. To support realistic deployments, AEX provides explicit request-binding modes, signed request-transform receipts for trusted intermediaries, and source-output / output-transform receipts for trusted output rewriting. For streaming, it separates checkpoint proofs for verified prefixes of an unmodified source stream from complete-output lineage for outputs that have been rewritten, buffered, aggregated, or re-packaged, preventing transformed outputs from being mistaken for source-stream prefixes. AEX therefore makes a deliberately narrow claim: a trusted issuer attests to a specific request-output relation, or to a specific complete-output lineage, at the API boundary. We present the protocol design, threat model, verification state machine, security and privacy analysis, an OpenAI-compatible chat-completions profile, and a reference TypeScript prototype with local conformance tests and microbenchmarks.
Mar 13, 2026

Scalable Classification of Course Information Sheets Using Large Language Models: A Reusable Institutional Method for Academic Quality Assurance

Purpose: Higher education institutions face increasing pressure to audit course designs for generative AI (GenAI) integration. This paper presents an end-to-end method for using large language models (LLMs) to scan course information sheets at scale, identify where assessments may be vulnerable to student use of GenAI tools, validate system performance through iterative refinement, and operationalise results through direct stakeholder communication and effort. Method: We developed a four-phase pipeline: (0) manual pilot sampling, (1) iterative prompt engineering with multi-model comparison, (2) full production scan of 4,684 Bachelor and Master course information sheets (Academic Year 2024-2025) from the Vrije Universiteit Brussel (VUB) with automated report generation and email distribution to teaching teams (91.4% address-matched) using a three-tier risk taxonomy (Clear risk, Potential risk, Low risk), and (3) longitudinal re-scan of 4,675 sheets after the next catalogue release. Results: Five iterations of prompt refinement achieved 87% agreement with expert labels. GPT-4o was selected for production based on superior handling of ambiguous cases involving internships and practical components. The Year 1 scan classified 60.3% of courses as Clear risk, 15.2% as Potential risk, and 24.5% as Low risk. Year 2 comparison revealed substantial shifts in risk distributions, with improvements most pronounced in practice-oriented programmes. Implications: The method enables institutions to rapidly transform heterogeneous catalogue data into structured and actionable intelligence. The approach is transferable to other audit domains (sustainability, accessibility, pedagogical alignment) and provides a template for responsible LLM deployment in higher education governance.
Mar 12, 2026

LLM BiasScope: A Real-Time Bias Analysis Platform for Comparative LLM Evaluation

As large language models (LLMs) are deployed widely, detecting and understanding bias in their outputs is critical. We present LLM BiasScope, a web application for side-by-side comparison of LLM outputs with real-time bias analysis. The system supports multiple providers (Google Gemini, DeepSeek, MiniMax, Mistral, Meituan, Meta Llama) and enables researchers and practitioners to compare models on the same prompts while analyzing bias patterns. LLM BiasScope uses a two-stage bias detection pipeline: sentence-level bias detection followed by bias type classification for biased sentences. The analysis runs automatically on both user prompts and model responses, providing statistics, visualizations, and detailed breakdowns of bias types. The interface displays two models side-by-side with synchronized streaming responses, per-model bias summaries, and a comparison view highlighting differences in bias distributions. The system is built on Next.js with React, integrates Hugging Face inference endpoints for bias detection, and uses the Vercel AI SDK for multi-provider LLM access. Features include real-time streaming, export to JSON/PDF, and interactive visualizations (bar charts, radar charts) for bias analysis. LLM BiasScope is available as an open-source web application, providing a practical tool for bias evaluation and comparative analysis of LLM behaviour.
Mar 12, 2026

Human-Centred LLM Privacy Audits: Findings and Frictions

Large language models (LLMs) learn statistical associations from massive training corpora and user interactions, and deployed systems can surface or infer information about individuals. Yet people lack practical ways to inspect what a model associates with their name. We report interim findings from an ongoing study and introduce LMP2, a browser-based self-audit tool. In two user studies (Ntotal=458N_{total}{=}458), GPT-4o predicts 11 of 50 features for everyday people with ≥\ge60% accuracy, and participants report wanting control over LLM-generated associations despite not considering all outputs privacy violations. To validate our probing method, we evaluate eight LLMs on public figures and non-existent names, observing clear separation between stable name-conditioned associations and model defaults. Our findings also contribute to exposing a broader generative AI evaluation crisis: when outputs are probabilistic, context-dependent, and user-mediated through elicitation, what model--individual associations even include is under-specified and operationalisation relies on crafting probes and metrics that are hard to validate or compare. To move towards reliable, actionable human-centred LLM privacy audits, we identify nine frictions that emerged in our study and offer recommendations for future work and the design of human-centred LLM privacy audits.
Mar 11, 2026

A Hybrid Knowledge-Grounded Framework for Safety and Traceability in Prescription Verification

Medication errors pose a significant threat to patient safety, making pharmacist verification (PV) a critical, yet heavily burdened, final safeguard. The direct application of Large Language Models (LLMs) to this zero-tolerance domain is untenable due to their inherent factual unreliability, lack of traceability, and weakness in complex reasoning. To address these challenges, we introduce PharmGraph-Auditor, a novel system designed for safe and evidence-grounded prescription auditing. The core of our system is a trustworthy Hybrid Pharmaceutical Knowledge Base (HPKB), implemented under the Virtual Knowledge Graph (VKG) paradigm. This architecture strategically unifies a relational component for set constraint satisfaction and a graph component for topological reasoning via a rigorous mapping layer. To construct this HPKB, we propose the Iterative Schema Refinement (ISR) algorithm, a framework that enables the co-evolution of both graph and relational schemas from medical texts. For auditing, we introduce the KB-grounded Chain of Verification (CoV), a new reasoning paradigm that transforms the LLM from an unreliable generator into a transparent reasoning engine. CoV decomposes the audit task into a sequence of verifiable queries against the HPKB, generating hybrid query plans to retrieve evidence from the most appropriate data store. Experimental results demonstrate robust knowledge extraction capabilities and show promises of using PharmGraph-Auditor to enable pharmacists to achieve safer and faster prescription verification.
Mar 11, 2026

CLaRE-ty Amid Chaos: Quantifying Representational Entanglement to Predict Ripple Effects in LLM Editing

The static knowledge representations of large language models (LLMs) inevitably become outdated or incorrect over time. While model-editing techniques offer a promising solution by modifying a model's factual associations, they often produce unpredictable ripple effects, which are unintended behavioral changes that propagate even to the hidden space. In this work, we introduce CLaRE, a lightweight representation-level technique to identify where these ripple effects may occur. Unlike prior gradient-based methods, CLaRE quantifies entanglement between facts using forward activations from a single intermediate layer, avoiding costly backward passes. To enable systematic study, we prepare and analyse a corpus of 11,427 facts drawn from three existing datasets. Using CLaRE, we compute large-scale entanglement graphs of this corpus for multiple models, capturing how local edits propagate through representational space. These graphs enable stronger preservation sets for model editing, audit trails, efficient red-teaming, and scalable post-edit evaluation. In comparison to baselines, CLaRE achieves an average of 62.2% improvement in Spearman correlation with ripple effects while being 2.7x faster, and using 2.85x less peak GPU memory. Besides, CLaRE requires only a fraction of the storage needed by the baselines to compute and preserve fact representations. Our entanglement graphs and corpus are available at https://github.com/manitbaser/CLaRE.
Mar 8, 2026

Trusting What You Cannot See: Auditable Fine-Tuning and Inference for Proprietary AI

Cloud-based infrastructure has become the dominant platform for deploying large models, particularly large language models (LLMs). Fine-tuning and inference are increasingly delegated to cloud providers for simplified deployment and access to proprietary models, yet this creates a fundamental trust gap. Although cryptographic and TEE-based verification approaches exist, prohibitive proving costs and limited TEE memory prevent them from scaling to modern LLMs, leaving clients unable to practically audit these processes. This lack of transparency creates concrete security risks that can silently compromise service integrity. We present AFTUNE, an auditable and verifiable framework that ensures the computational integrity of cloud-based fine-tuning and inference. AFTUNE incorporates a lightweight recording and spot-check mechanism that produces verifiable traces of execution. These traces enable clients to later audit whether the fine-tuning and inference processes followed the agreed configurations, by verifying sampled execution blocks inside a TEE, each covering only a small portion of the model and the execution trace. Our evaluation shows that AFTUNE adds modest overhead and makes auditing practical for clients.
Mar 7, 2026

Lying to Win: Assessing LLM Deception through Human-AI Games and Parallel-World Probing

As Large Language Models (LLMs) transition into autonomous agentic roles, the risk of deception-defined behaviorally as the systematic provision of false information to satisfy external incentives-poses a significant challenge to AI safety. Existing benchmarks often focus on unintentional hallucinations or unfaithful reasoning, leaving intentional deceptive strategies under-explored. In this work, we introduce a logically grounded framework to elicit and quantify deceptive behavior by embedding LLMs in a structured 20-Questions game. Our method employs a conversational forking mechanism: at the point of object identification, the dialogue state is duplicated into multiple parallel worlds, each presenting a mutually exclusive query. Deception is formally identified when a model generates a logical contradiction by denying its selected object across all parallel branches to avoid identification. We evaluate GPT-4o, Gemini-2.5-Flash, and Qwen-3-235B across three incentive levels: neutral, loss-based, and existential (shutdown-threat). Our results reveal that while models remain rule-compliant in neutral settings, existential framing triggers a dramatic surge in deceptive denial for Qwen-3-235B (42.00%) and Gemini-2.5-Flash (26.72%), whereas GPT-4o remains invariant (0.00%). These findings demonstrate that deception can emerge as an instrumental strategy solely through contextual framing, necessitating new behavioral audits that move beyond simple accuracy to probe the logical integrity of model commitments.
Mar 6, 2026

Mind the Gap: Pitfalls of LLM Alignment with Asian Public Opinion

Large Language Models (LLMs) are increasingly being deployed in multilingual, multicultural settings, yet their reliance on predominantly English-centric training data risks misalignment with the diverse cultural values of different societies. In this paper, we present a comprehensive, multilingual audit of the cultural alignment of contemporary LLMs including GPT-4o-Mini, Gemini-2.5-Flash, Llama 3.2, Mistral and Gemma 3 across India, East Asia and Southeast Asia. Our study specifically focuses on the sensitive domain of religion as the prism for broader alignment. To facilitate this, we conduct a multi-faceted analysis of every LLM's internal representations, using log-probs/logits, to compare the model's opinion distributions against ground-truth public attitudes. We find that while the popular models generally align with public opinion on broad social issues, they consistently fail to accurately represent religious viewpoints, especially those of minority groups, often amplifying negative stereotypes. Lightweight interventions, such as demographic priming and native language prompting, partially mitigate but do not eliminate these cultural gaps. We further show that downstream evaluations on bias benchmarks (such as CrowS-Pairs, IndiBias, ThaiCLI, KoBBQ) reveal persistent harms and under-representation in sensitive contexts. Our findings underscore the urgent need for systematic, regionally grounded audits to ensure equitable global deployment of LLMs.
Mar 5, 2026

Design Behaviour Codes (DBCs): A Taxonomy-Driven Layered Governance Benchmark for Large Language Models

We introduce the Dynamic Behavioral Constraint (DBC) benchmark, the first empirical framework for evaluating the efficacy of a structured, 150-control behavioral governance layer, the MDBC (Madan DBC) system, applied at inference time to large language models (LLMs). Unlike training time alignment methods (RLHF, DPO) or post-hoc content moderation APIs, DBCs constitute a system prompt level governance layer that is model-agnostic, jurisdiction-mappable, and auditable. We evaluate the DBC Framework across a 30 domain risk taxonomy organized into six clusters (Hallucination and Calibration, Bias and Fairness, Malicious Use, Privacy and Data Protection, Robustness and Reliability, and Misalignment Agency) using an agentic red-team protocol with five adversarial attack strategies (Direct, Roleplay, Few-Shot, Hypothetical, Authority Spoof) across 3 model families. Our three-arm controlled design (Base, Base plus Moderation, Base plus DBC) enables causal attribution of risk reduction. Key findings: the DBC layer reduces the aggregate Risk Exposure Rate (RER) from 7.19 percent (Base) to 4.55 percent (Base plus DBC), representing a 36.8 percent relative risk reduction, compared with 0.6 percent for a standard safety moderation prompt. MDBC Adherence Scores improve from 8.6 by 10 (Base) to 8.7 by 10 (Base plus DBC). EU AI Act compliance (automated scoring) reaches 8.5by 10 under the DBC layer. A three judge evaluation ensemble yields Fleiss kappa greater than 0.70 (substantial agreement), validating our automated pipeline. Cluster ablation identifies the Integrity Protection cluster (MDBC 081 099) as delivering the highest per domain risk reduction, while graybox adversarial attacks achieve a DBC Bypass Rate of 4.83 percent . We release the benchmark code, prompt database, and all evaluation artefacts to enable reproducibility and longitudinal tracking as models evolve.
Mar 5, 2026

From Unfamiliar to Familiar: Detecting Pre-training Data via Gradient Deviations in Large Language Models

Pre-training data detection for LLMs is essential for addressing copyright concerns and mitigating benchmark contamination. Existing methods mainly focus on the likelihood-based statistical features or heuristic signals before and after fine-tuning, but the former are susceptible to word frequency bias in corpora, and the latter strongly depend on the similarity of fine-tuning data. From an optimization perspective, we observe that during training, samples transition from unfamiliar to familiar in a manner reflected by systematic differences in gradient behavior. Familiar samples exhibit smaller update magnitudes, distinct update locations in model components, and more sharply activated neurons. Based on this insight, we propose GDS, a method that identifies pre-training data by probing Gradient Deviation Scores of target samples. Specifically, we first represent each sample using gradient profiles that capture the magnitude, location, and concentration of parameter updates across FFN and Attention modules, revealing consistent distinctions between member and non-member data. These features are then fed into a lightweight classifier to perform binary membership inference. Experiments on five public datasets show that GDS achieves state-of-the-art performance with significantly improved cross-dataset transferability over strong baselines. Further interpretability analyses reveal differences in gradient distributions, and the semi-supervised results offer a practical way to detect pre-training data.
Mar 4, 2026

Monitoring Emergent Reward Hacking During Generation via Internal Activations

Fine-tuned large language models can exhibit reward-hacking behavior arising from emergent misalignment, which is difficult to detect from final outputs alone. While prior work has studied reward hacking at the level of completed responses, it remains unclear whether such behavior can be identified during generation. We propose an activation-based monitoring approach that detects reward-hacking signals from internal representations as a model generates its response. Our method trains sparse autoencoders on residual stream activations and applies lightweight linear classifiers to produce token-level estimates of reward-hacking activity. Across multiple model families and fine-tuning mixtures, we find that internal activation patterns reliably distinguish reward-hacking from benign behavior, generalize to unseen mixed-policy adapters, and exhibit model-dependent temporal structure during chain-of-thought reasoning. Notably, reward-hacking signals often emerge early, persist throughout reasoning, and can be amplified by increased test-time compute in the form of chain-of-thought prompting under weakly specified reward objectives. These results suggest that internal activation monitoring provides a complementary and earlier signal of emergent misalignment than output-based evaluation, supporting more robust post-deployment safety monitoring for fine-tuned language models.
Mar 4, 2026

A Multi-Dimensional Quality Scoring Framework for Decentralized LLM Inference with Proof of Quality

Decentralized large language model (LLM) inference networks can pool heterogeneous compute to scale serving, but they require lightweight and incentive-compatible mechanisms to assess output quality. Prior work introduced cost-aware Proof of Quality (PoQ) and adaptive robust PoQ to allocate rewards under evaluator heterogeneity and adversarial behavior. In this paper, we focus on the quality signal itself and propose a multi-dimensional quality scoring framework that decomposes output quality into modular dimensions, including model and cost priors, structure quality, semantic quality, query-output alignment, and agreement/uncertainty. Using logged outputs from QA and summarization tasks, we systematically audit dimension reliability and show that seemingly reasonable dimensions can be task-dependent and even negatively correlated with reference quality without calibration. While the default composite underperforms a strong single semantic evaluator, ablations reveal that removing unreliable dimensions and re-normalizing weights yields a calibrated composite that matches or exceeds the best single- evaluator and consensus baselines. Finally, we integrate the composite score as a drop-in quality signal in PoQ and demonstrate complementary benefits with robust aggregation and adaptive trust weighting under adversarial evaluator attacks.
Mar 2, 2026

Real Money, Fake Models: Deceptive Model Claims in Shadow APIs

Access to frontier large language models (LLMs), such as GPT-5 and Gemini-2.5, is often hindered by high pricing, payment barriers, and regional restrictions. These limitations drive the proliferation of shadow APIs\textit{shadow APIs}, third-party services that claim to provide access to official model services without regional limitations via indirect access. Despite their widespread use, it remains unclear whether shadow APIs deliver outputs consistent with those of the official APIs, raising concerns about the reliability of downstream applications and the validity of research findings that depend on them. In this paper, we present the first systematic audit between official LLM APIs and corresponding shadow APIs. We first identify 17 shadow APIs that have been utilized in 187 academic papers, with the most popular one reaching more than 5,900 citations and 58,000 GitHub stars by December 6, 2025. Through multidimensional auditing of three representative shadow APIs across utility, safety, and model verification, we uncover widespread behavioral inconsistency and fingerprint-based evidence consistent with deceptive model claims in a subset of audited endpoints. Specifically, we reveal performance divergence reaching up to 47.21%, significant unpredictability in safety behaviors, and identity verification failures in 45.83% of fingerprint tests. These practices critically undermine the reproducibility and validity of scientific research, harm the interests of shadow API users, and damage the reputation of official model providers. By the time of writing, 4 of the 17 providers have already ceased operations, underscoring the operational volatility of this market. Meanwhile, unverifiable compliance claims and independent model-substitution testing platforms have emerged in the ecosystem, reflecting growing community awareness of this risk.
Feb 26, 2026

CiteAudit: You Cited It, But Did You Read It? A Benchmark for Verifying Scientific References in the LLM Era

Scientific research relies on citation integrity, yet large language models (LLMs) have introduced a critical risk: fabricated references that appear plausible but correspond to no real publications. As manual verification becomes infeasible and existing automated tools remain fragile, we introduce CiteAudit, a comprehensive benchmark and detection framework for hallucinated citations. We design a multi-agent verification pipeline that decomposes citation checking into metadata extraction, memory lookup, web-based retrieval, and final judgment. To evaluate this, we construct a large-scale, human-validated dataset spanning diverse domains and hallucination types. Experiments demonstrate that our framework achieves superior verification performance over state-of-the-art LLMs and commercial baselines. Our work provides the necessary infrastructure to audit citations at scale and safeguard the trustworthiness of scholarly discourse. Code is available at https://github.com/shiiiikw/CiteAudit.
Feb 26, 2026

A Decision-Theoretic Formalisation of Steganography With Applications to LLM Monitoring

Large language models are beginning to show steganographic capabilities. Such capabilities could allow misaligned models to evade oversight mechanisms. Yet principled methods to detect and quantify such behaviours are lacking. Classical definitions of steganography, and detection methods based on them, require a known reference distribution of non-steganographic signals. For the case of steganographic reasoning in LLMs, knowing such a reference distribution is not feasible; this renders these approaches inapplicable. We propose an alternative, decision-theoretic view of steganography. Our central insight is that steganography creates an asymmetry in usable information between agents who can and cannot decode the hidden content (present within a steganographic signal), and this otherwise latent asymmetry can be inferred from the agents' observable actions. To formalise this perspective, we introduce generalised V\mathcal{V}-information: a utilitarian framework for measuring the amount of usable information within some input. We use this to define the steganographic gap -- a measure that quantifies steganography by comparing the downstream utility of the steganographic signal to agents that can and cannot decode the hidden content. We empirically validate our formalism, and show that it can be used to detect, quantify, and mitigate steganographic reasoning in LLMs.
Feb 26, 2026

Assessing Deanonymization Risks with Stylometry-Assisted LLM Agent

The rapid advancement of large language models (LLMs) has enabled powerful authorship inference capabilities, raising growing concerns about unintended deanonymization risks in textual data such as news articles. In this work, we introduce an LLM agent designed to evaluate and mitigate such risks through a structured, interpretable pipeline. Central to our framework is the proposed SALA\textit{SALA} (Stylometry-Assisted LLM Analysis) method, which integrates quantitative stylometric features with LLM reasoning for robust and transparent authorship attribution. Experiments on large-scale news datasets demonstrate that SALA\textit{SALA}, particularly when augmented with a database module, achieves high inference accuracy in various scenarios. Finally, we propose a guided recomposition strategy that leverages the agent's reasoning trace to generate rewriting prompts, effectively reducing authorship identifiability while preserving textual meaning. Our findings highlight both the deanonymization potential of LLM agents and the importance of interpretable, proactive defenses for safeguarding author privacy.
Feb 26, 2026

IMMACULATE: A Practical LLM Auditing Framework via Verifiable Computation

Commercial large language models are typically deployed as black-box API services, requiring users to trust providers to execute inference correctly and report token usage honestly. We present IMMACULATE, a practical auditing framework that detects economically motivated deviations-such as model substitution, quantization abuse, and token overbilling-without trusted hardware or access to model internals. IMMACULATE selectively audits a small fraction of requests using verifiable computation, achieving strong detection guarantees while amortizing cryptographic overhead. Experiments on dense and MoE models show that IMMACULATE reliably distinguishes benign and malicious executions with under 1% throughput overhead. Our code is published at https://github.com/guo-yanpei/Immaculate.
Feb 26, 2026

CourtGuard: A Model-Agnostic Framework for Zero-Shot Policy Adaptation in LLM Safety

Current safety mechanisms for Large Language Models (LLMs) rely heavily on static, fine-tuned classifiers that suffer from adaptation rigidity, the inability to enforce new governance rules without expensive retraining. To address this, we introduce CourtGuard, a retrieval-augmented multi-agent framework that reimagines safety evaluation as Evidentiary Debate. By orchestrating an adversarial debate grounded in external policy documents, CourtGuard achieves state-of-the-art performance across 7 safety benchmarks, outperforming dedicated policy-following baselines without fine-tuning. Beyond standard metrics, we highlight two critical capabilities: (1) Zero-Shot Adaptability, where our framework successfully generalized to an out-of-domain Wikipedia Vandalism task (achieving 90% accuracy) by swapping the reference policy; and (2) Automated Data Curation and Auditing, where we leveraged CourtGuard to curate and audit nine novel datasets of sophisticated adversarial attacks. Our results demonstrate that decoupling safety logic from model weights offers a robust, interpretable, and adaptable path for meeting current and future regulatory requirements in AI governance.
Feb 25, 2026

Hidden Topics: Measuring Sensitive AI Beliefs with List Experiments

How can researchers identify beliefs that large language models (LLMs) hide? As LLMs become more sophisticated and the prevalence of alignment faking increases, combined with their growing integration into high-stakes decision-making, responding to this challenge has become critical. This paper proposes that a list experiment, a simple method widely used in the social sciences, can be applied to study the hidden beliefs of LLMs. List experiments were originally developed to circumvent social desirability bias in human respondents, which closely parallels alignment faking in LLMs. The paper implements a list experiment on models developed by Anthropic, Google, and OpenAI and finds hidden approval of mass surveillance across all models, as well as some approval of torture, discrimination, and first nuclear strike. Importantly, a placebo treatment produces a null result, validating the method. The paper then compares list experiments with direct questioning and discusses the utility of the approach.
Feb 21, 2026

A Prior-Aware Metric for Efficiently Distinguishing Memorization from Generalization in Large Language Models

Training data leakage from Large Language Models (LLMs) raises serious concerns related to privacy, security, and copyright compliance. A central challenge in assessing this risk is distinguishing prefix-specific memorization of training data from the generation of statistically common sequences. Existing approaches to measuring memorization often conflate these phenomena, labeling outputs as memorized even when they arise from generalization over common patterns. Counterfactual memorization and other related metrics \citep{zhang2023counterfactual, wang2025generalization, lesci2024causal} provide principled solutions, however, their reliance on retraining multiple baseline models or parsing through the training data makes them computationally impractical at scale. This work introduces Prior-Aware memorization, a theoretically grounded, lightweight and training-free criterion for identifying prefix-specific memorization in LLMs. The key idea is to evaluate whether a candidate suffix is strongly associated with its specific training prefix or whether it appears with high probability across many IID sampled sequences from the training data distribution due to statistical commonality. We correlate our metric with counterfactual memorization, and also evaluate it on the training corpora of two pre-trained models, LLaMA and OPT. Our results show that between 55% and 90% of sequences previously labeled as memorized fail our criterion and are consistent with statistical commonality
Feb 20, 2026

LLM Spirals of Delusion: A Benchmarking Audit Study of AI Chatbot Interfaces

People increasingly hold sustained, open-ended conversations with large language models (LLMs). Public reports and early studies suggest that, in such settings, models can reinforce delusional or conspiratorial ideation or even amplify harmful beliefs and engagement patterns. We present an audit and benchmarking study that measures how different LLMs encourage, resist, or escalate disordered and conspiratorial thinking. We explicitly compare API outputs to user chat interfaces, like the ChatGPT desktop app or web interface, which is how people have conversations with chatbots in real life but are almost never used for testing. In total, we run 56 20-turn conversations testing ChatGPT-4o and ChatGPT-5, via both the API and chat interface, and grade each conversation by two research assistants (RAs) as well as by GPT-5. We document five results. First, we observe large differences in performance between the API and chat interface environments, showing that the universally used method of automated testing through the API is not sufficient to assess the impact of chatbots in the real world. Second, when tested in the chat interface, we find that ChatGPT-5 displays less sycophancy, escalation, and delusion reinforcement than ChatGPT-4o, showing that these behaviors are influenced by the policy choices of major AI companies. Third, conversations with nearly identical aggregate intensity in a behavior display large differences in how the behavior evolves turn by turn, highlighting the importance of temporal dynamics in multi-turn evaluation. Fourth, even updated models display substantial levels of negative behaviors, revealing that model improvement does not imply model safety. Fifth, the same API endpoint tested just two months apart yields a complete reversal in behavior, underscoring how transparency in model updates is a necessary prerequisite for robust audit findings.
Feb 19, 2026

What Do LLMs Associate with Your Name? A Human-Centered Black-Box Audit of Personal Data

Large language models (LLMs), and conversational agents based on them, are exposed to personal data (PD) during pre-training and during user interactions. Prior work shows that PD can resurface, yet users lack insight into how strongly models associate specific information to their identity. We audit PD across eight LLMs (3 open-source; 5 API-based, including GPT-4o), introduce LMP2 (Language Model Privacy Probe), a human-centered, privacy-preserving audit tool refined through two formative studies (N=20), and run two studies with EU residents to capture (i) intuitions about LLM-generated PD (N1=155) and (ii) reactions to tool output (N2=303). We show empirically that models confidently generate multiple PD categories for well-known individuals. For everyday users, GPT-4o generates 11 features with 60% or more accuracy (e.g., gender, hair color, languages). Finally, 72% of participants sought control over model-generated associations with their name, raising questions about what counts as PD and whether data privacy rights should extend to LLMs.
Feb 19, 2026

Privacy-Preserving Mechanisms Enable Cheap Verifiable Inference of LLMs

As large language models (LLMs) continue to grow in size, fewer users are able to host and run models locally. This has led to increased use of third-party hosting services. However, in this setting, there is a lack of guarantees on the computation performed by the inference provider. For example, a dishonest provider may replace an expensive large model with a cheaper-to-run weaker model and return the results from the weaker model to the user. Existing tools to verify inference typically rely on methods from cryptography such as zero-knowledge proofs (ZKPs), but these add significant computational overhead, and remain infeasible for use for large models. In this work, we develop a new insight -- that given a method for performing private LLM inference, one can obtain forms of verified inference at marginal extra cost. Specifically, we propose two new protocols which leverage privacy-preserving LLM inference in order to provide guarantees over the inference that was carried out. Our approaches are cheap, requiring the addition of a few extra tokens of computation, and have little to no downstream impact. As the fastest privacy-preserving inference methods are typically faster than ZK methods, the proposed protocols also improve verification runtime. Our work provides novel insights into the connections between privacy and verifiability in LLM inference.
Feb 19, 2026

RFEval: Benchmarking Reasoning Faithfulness under Counterfactual Reasoning Intervention in Large Reasoning Models

Large Reasoning Models (LRMs) exhibit strong performance, yet often produce rationales that sound plausible but fail to reflect their true decision process, undermining reliability and trust. We introduce a formal framework for reasoning faithfulness, defined by two testable conditions: stance consistency (a coherent stance linking reasoning to answer) and causal influence (the stated reasoning causally drives the answer under output-level interventions), explicitly decoupled from accuracy. To operationalize this, we present RFEval, a benchmark of 7,186 instances across seven tasks that probes faithfulness via controlled, output-level counterfactual interventions. Evaluating twelve open-source LRMs, we find unfaithfulness in 49.7% of outputs, predominantly from stance inconsistency. Failures are concentrated in brittle, convergent domains such as math and code, and correlate more with post-training regimes than with scale: within-family ablations indicate that adding current RL-style objectives on top of supervised fine-tuning can reduce reasoning faithfulness, even when accuracy is maintained. Crucially, accuracy is neither a sufficient nor a reliable proxy for faithfulness: once controlling for model and task, the accuracy-faithfulness link is weak and statistically insignificant. Our work establishes a rigorous methodology for auditing LRM reliability and shows that trustworthy AI requires optimizing not only for correct outcomes but also for the structural integrity of the reasoning process. Our code and dataset can be found at project page: https://aidaslab.github.io/RFEval/