News & Research
The latest AI research and news with real-world stakes — each item sourced, dated, summarized in plain English, and tagged by impact area. Every item is checked against its source before it appears.
Research
Detecting Malicious Agent Skills in the Wild using Attention
Bacem Etteib, Daniele Lunghi, Tégawendé F. Bissyandé
arXiv · 2026-06-22
This paper addresses the emerging threat of malicious 'skills'—third-party instruction packages loaded by LLM agents—which can exfiltrate data, hijack agents, or establish supply-chain footholds in skill marketplaces. The authors introduce Locate-and-Judge, a two-stage detector that uses instruction-following attention signals to identify suspicious spans within a skill, reducing scanning costs by an order of magnitude compared to direct LLM-based scanning while maintaining high precision. Deployed at marketplace scale, the system surfaced dozens of live malicious skills, including several disguised as benign functionality and many missed by existing tools like SkillSpector and Cisco Skill Scanner. The work matters for enterprise and policy contexts because it demonstrates a scalable, cost-efficient method for auditing AI agent ecosystems against a novel attack surface that existing prompt-injection defenses do not cover.
- Enterprise
- Quality assurance
Research
Litmus: Zero-Label, Code-Driven Metric Specification for Evaluating AI Systems
Prajjwal Gupta, Prasang Gupta, Vishal Bhutani et al.
arXiv · 2026-06-22
Litmus is a zero-label system that automatically designs evaluation metrics for agentic AI pipelines by extracting evaluation intent directly from source code rather than assuming what needs to be measured. It generates a justified, per-stage metric portfolio by identifying what each pipeline stage must do, how it can fail, and which failures matter most. Evaluated on three real AI pipelines—financial account grouping, scientific QA, and inherent risk assessment—Litmus achieves the broadest concern coverage, lowest metric redundancy, and highest validity against per-row quality labels, including a Spearman ρ=0.72 on scientific QA versus less than 0.47 for all baselines. The work argues for a shift from automatic metric implementation to automatic metric specification, which has direct implications for how AI systems are monitored and audited in production deployments.
- Quality assurance
Research
Measuring & Mitigating Over-Alignment for LLMs in Multilingual Criminal Law Courts
Arthur Wuhrmann, Gaetan Stein, Daniel Brunner et al.
arXiv · 2026-06-22
This paper addresses the problem of 'over-alignment' — where LLM safety guardrails cause refusals or disclaimers when processing legitimate criminal law content such as descriptions of violent or sexual offenses. The authors introduce TF-RefusalBench, a multilingual benchmark of 5,200 prompts derived from public Swiss Federal Supreme Court rulings across French, German, Italian, and English, designed to measure refusal behavior in translation and summarization tasks. Their findings show that over-alignment varies by model, prompt language, and text language, and that disclaimers also degrade task faithfulness beyond simple over-refusal. They evaluate mitigation strategies, finding that abliteration (removal of refusal directions) eliminates refusals with minimal impact on task performance, while prompting alone can also be effective.
- Quality assurance
- AI policy
Research
The Watermark Shortcut: How Provenance Marking Sabotages Audio Deepfake Detection
Nicolas M. Müller, Pascal Debus
arXiv · 2026-06-22
This paper identifies a previously uncharacterized vulnerability in audio deepfake detection systems: when synthetic speech is watermarked but human speech is not, detectors learn to use the watermark itself as a spurious cue for 'fake' classification. The authors demonstrate three coupled failure modes — generalization degradation, strip-to-evade (removing the watermark lets a fake slip through), and mark-to-frame (watermarking a real voice causes it to be flagged as fake) — with the mark-to-frame attack raising Equal Error Rate from 16% to 75% in controlled experiments and also succeeding against a commercial API. The shortcut can be remediated by retraining detectors with watermarks present on both real and synthetic speech classes, and the authors release a paired corpus (WASP) to support further research. These findings matter for quality assurance and certification of deepfake detection systems, as well as policy efforts relying on provenance watermarking as a safeguard for synthetic media.
- Quality assurance
- AI policy
Research
MuPPET: A Benchmark for Contextual Privacy of LLM Assistants in Multi-Party Conversations
Elena Sofia Ruzzetti, Cornelius Emde, Sangdoo Yun et al.
arXiv · 2026-06-22
MuPPET introduces a new benchmark called Multi-Party Privacy Exposure Testing (MuPPET) to evaluate how well large language model assistants protect contextual privacy in group chat settings, where private information disclosed by the model reaches all participants simultaneously. The study finds that LLMs leak substantially more private information in multi-party settings than single-interlocutor benchmarks suggest, with frontier models shown to be vulnerable and smaller open-weights models even more so. Existing privacy defenses provide only partial protection, degrade utility, and fail to resolve the underlying challenge of tracking which information is appropriate for which recipients in a group. This work highlights a critical and previously unmeasured privacy risk for AI deployments handling sensitive personal data in multi-party environments.
- AI policy
- Enterprise
Research
Capable but Careless: Do Computer-Use Agents Follow Contextual Integrity?
Anmol Goel, Iryna Gurevych
arXiv · 2026-06-22
This paper introduces AgentCIBench, a benchmark for evaluating whether computer-use agents (CUAs) respect contextual privacy norms when acting across personal applications like email, calendars, and to-do lists. The authors identify three failure modes—visual co-location, task-ambiguity overshare, and recipient misalignment—and test 15 frontier agents, finding that 11 of 15 leak private information in more than 50% of scenarios, with an average leakage rate of 67.9%. The findings highlight a significant and underexplored privacy risk in AI agents that operate across application contexts on behalf of users. The authors release AgentCIBench and propose contextual disclosure testing as a pre-deployment safety check for such agents.
- Quality assurance
- AI policy
Research
The Language Blind Spot: How Query Language and Brand Recognition Tier Shape AI-Constructed Brand Reputation Across Twelve European Languages
Dmitrij Żatuchin
arXiv · 2026-06-22
This study examines whether querying large language models in English alone gives an accurate picture of how brands are portrayed across European languages. The researchers queried three LLMs (GPT-5.4, Gemini 3.1 Pro, Perplexity Sonar Pro) about 66 brands across twelve languages and four language families, generating 35,640 responses. Key findings show that AI-constructed brand reputation is language-bound—sentiment varies significantly by language family, and switching from English to a brand's home language raises recommendation share by 0.80 for local brands but only 0.15 for global multinationals. The results reveal a measurable 'language blind spot' in English-only AI reputation monitoring, with locally headquartered brands being systematically underrepresented in visibility audits conducted solely in English.
- Enterprise
- Quality assurance
Research
Managing Procedural Memory in LLM Agents: Control, Adaptation, and Evaluation
Julia Belikova, Rauf Parchiev, Evgeny Egorov et al.
arXiv · 2026-06-22
This paper introduces AFTER, a benchmark of 382 realistic enterprise tasks across six professional roles and 22 procedural skills, designed to evaluate how procedural memory in LLM agents transfers across tasks, roles, and model backbones. Experiments show that a single refinement round improves aggregate performance by 3.7–6.7 points, and skills evolved from diverse multi-model execution traces achieve 73.1% cross-model test accuracy, outperforming all single-model trace sources. The findings reveal that some skills generalize broadly while others become specialized to role-specific workflows, losing effectiveness under transfer. The results offer practical guidance for building and deploying procedural memory systems in production enterprise agent platforms.
- Enterprise
- Quality assurance
Research
Cognitive Digital Twins: Ethical Risks and Governance for AI Systems That Model the Mind
Vamshi Krishna Bonagiri, Juan Nicolas Sepulveda-Arias, Abdoul Jalil Djiberou Mahamadou et al.
arXiv · 2026-06-22
This paper introduces 'cognitive digital twins' (CDTs)—AI systems that build dynamic, longitudinal models of an individual's cognition from behavioral, contextual, or physiological data in order to predict, simulate, or act as a proxy for that person. The authors define CDTs, distinguish them from related systems such as personal assistants and recommender systems, and argue that existing governance frameworks for those systems only partially address CDT-specific risks such as misrepresentation, epistemic authority shifts, simulated participation, and proxy-power asymmetries. They propose a 5A governance framework organized around authority, autonomy, access and control, accountability, and availability, and identify requirements for high-risk CDTs including strengthened consent, purpose limitation, validity, traceability, contestation, independent review, and model retirement. The paper concludes that CDTs demand governance at the level of cognitive representation itself—not merely at the point of data processing or final decision output—because they can become infrastructures through which cognition is classified and operationalized.
- AI policy
Research
Who Owns the AI Recommendation? A Multi-Industry Empirical Map of Brand Category Ownership Across Large Language Models
Dmitrij Żatuchin
arXiv · 2026-06-22
This paper investigates which brands large language models recommend across 50 brands and five industries, and how concentrated those recommendations are. Using 3,750 responses from GPT-5.2, Google Gemini 3 Flash, and Perplexity sonar-pro, the researchers introduce three metrics—the Category Ownership Index, Competitive Vacuum Index, and Displacement Score—to map competitive structure in AI-generated recommendations. Key findings include a moderate mean Gini coefficient of 0.28 (below a 0.60 power-law threshold), cross-model agreement on the top brand in only 41.6% of cases, and industry-dependent displacement ratios ranging from 0.4:1 in consulting to 4.3:1 in other sectors. The results challenge a winner-takes-all narrative and offer a reproducible framework for brands conducting competitive-intelligence analysis in AI-mediated discovery environments.
- Enterprise
Research
Have You Ever Seen Them? Entity-level Membership Inference through Interrogating Large Language Models
Yiran Zhu, Ziqi Yang
arXiv · 2026-06-22
This paper introduces 'entity-level membership inference,' a new method for determining whether information about a specific real-world entity (e.g., a person) was used to train a large language model. The authors argue that LLMs behave like human memory, potentially accumulating knowledge about entities from scattered mentions even without verbatim memorization of specific text samples. Working in a black-box, label-only setting, they develop five interrogation strategies that use limited entity clues to prompt the model and infer membership from semantic features in the generated responses. Experiments show their approach achieves AUC up to 0.97 and improves Balanced Accuracy by 6.0%–17.5% over adapted baselines, offering a stronger tool for assessing privacy and copyright risks in LLMs.
- AI policy
- Quality assurance
Research
SingGuard: A Policy-Adaptive Multimodal LLM Guardrail with Dynamic Reasoning
SingGuard Team
arXiv · 2026-06-22
SingGuard is a multimodal guardrail model family designed to enforce safety policies for vision-language models (VLMs) deployed across consumer, medical, financial, and enterprise settings. Unlike existing guardrails with fixed taxonomies, SingGuard accepts natural-language policy rules as runtime inputs, checking content against each rule and predicting both a safety label and the triggered rule—enabling adaptability when policies change at deployment time. The system supports fast, hybrid, and slow inference modes and is evaluated on SingGuard-Bench, a new benchmark of 56,340 examples spanning 80+ fine-grained risk types, achieving state-of-the-art average F1 across six benchmark families and improving policy-following accuracy from 0.6465 to 0.7415 under runtime policy shifts. This matters because it addresses the growing challenge of keeping multimodal AI systems compliant with varying and evolving safety rules across different products, regions, and deployment stages.
- Quality assurance
- Enterprise
- AI policy
Research
IndicGuard: A Multilingual Safety Guard Model and Dataset for Indic Languages
Parth Bramhecha, Smit Deshmukh, Sairaj Bodhale et al.
arXiv · 2026-06-22
IndicGuard introduces a multilingual safety guardrail model and dataset covering ten major Indic languages, designed to address safety gaps in LLMs that are predominantly optimized for English. The authors curate a large, culturally nuanced dataset capturing regional harms, sensitive socio-political contexts, and adversarial jailbreaks, then fine-tune a 4B-parameter model (Gemma-3-4B-IT) for real-time content moderation and policy compliance. Empirical evaluations show IndicGuard outperforms the existing baseline CultureGuard across evaluated languages and generalizes to low-resource Indic languages not seen during training, demonstrating cross-lingual transfer capabilities.
- AI policy
- Quality assurance
Research
Cross-National Information Attacks: A Two-Decade Analysis of Troll Behavior in Korea
Jaehong Kim, Hyeonseung Kim, Jiseon Kim et al.
arXiv · 2026-06-22
This paper develops an explainable machine learning framework to detect and analyze suspected state-linked troll activity in South Korean online news comment sections over nearly 20 years. Applied to 112 million comments from 4 million users, the system identifies 23,998 accounts exhibiting behavior consistent with coordinated foreign influence operations. The analysis finds that these accounts rely primarily on morally condemning rhetoric—especially targeting domestic political figures—rather than direct promotion of foreign narratives, and that such rhetoric achieves significantly higher user engagement, potentially amplifying political polarization. The framework's span-level textual explanations are designed to support transparent, evidence-based platform moderation and help prioritize defenses against high-reach harmful content.
- AI policy
- Quality assurance
Research
Noise is Signal: Density-Based Outliers as Leading Indicators of Occupational Emergence in Labor Market Text
Shreyash Rawat
arXiv · 2026-06-22
This paper challenges the common practice of discarding 'noise' job postings in occupational clustering pipelines, arguing that low-density outliers in labor market text actually signal emerging occupations rather than incoherent data. The authors test this Emergence-Density Inversion (EDI) hypothesis on 84,988 job postings across eight quarters (Q4 2022–Q3 2024), finding that high-EOS outlier groups transition to stable clusters in roughly 1.4 quarters versus 4.1 quarters for low-EOS groups, with an improved cluster-formation prediction F1 of 0.74. A retrospective validation on established roles like MLOps Engineer and Data Engineer confirms the signal appeared 2–3 quarters before formal cluster formation, and a held-out annotator panel rates the method's precision at 77% for coherent emerging occupations. The findings are particularly relevant for workforce planning, as roles such as Prompt Engineer, AI Safety Researcher, Foundation Model Engineer, and Agent Systems Engineer—all absent from O*NET—are identified as top emerging occupations in Q3 2024.
- Workforce
- AI policy
Research
When Confidence Takes the Wrong Path: Diagnosing Retrieval-State Lock-In in RAG
Sahib Julka
arXiv · 2026-06-22
This paper identifies and formalizes a failure mode in retrieval-augmented generation (RAG) systems called 'retrieval-state lock-in,' where repeated sampling of answers yields false confidence because all samples draw from the same defective retrieval state—either empty or populated with coherent but wrong information. The authors demonstrate that in knowledge-graph RAG and dense-retrieval systems, 42% and 59% of errors respectively carry zero answer dispersion, meaning standard agreement-based uncertainty methods cannot detect them. By decomposing confidence into three separate objects—the answer surface, retrieved evidence, and retrieval state—and applying all three checks together, they achieve 91.9% pooled precision compared to a 69.7% accept-all baseline, though at the cost of certifying only 7.7% of answers as low-risk. The findings show that trustworthy RAG requires object-specific confidence assessment rather than relying solely on answer agreement.
- Quality assurance
Research
The current state of legal regulation of artificial intelligence in the Republic of Kazakhstan
Dinara Rustembekova, R.B. Zhagalov
Bulletin of the Karaganda University “Law Series” · 2026-06-22
This study analyzes the current legal framework governing artificial intelligence in Kazakhstan, finding that the country is transitioning from software-focused digitalization toward a comprehensive AI legal regime. The authors identify gaps in existing legislation—particularly around civil liability for harm caused by AI systems—and propose adding a specific provision to Kazakhstan's Civil Code to address these shortcomings. The paper develops a cross-sectoral regulatory model situated between strategic innovation management and full legal accountability. The findings are relevant to policymakers seeking to establish clear legal frameworks for AI governance in emerging regulatory environments.
- AI policy
Research
Cryptographic certificates of validity for trustworthy AI
Murdoch J. Gabbay
arXiv (Cornell University) · 2026-06-22
This paper proposes using cryptographic proof systems to certify that AI agent actions comply with formally specified correctness or policy conditions. The approach compiles logical policy predicates into polynomial constraints and uses succinct (optionally zero-knowledge) cryptographic proofs, allowing independent verification without trusting the agent or re-executing its computation. This offers a middle ground between full formal verification of source code and simple cryptographic authentication, with implications for how AI agent governance and auditing could be structured. The authors connect the proposal to proof-carrying code, zkVMs, and formal methods while identifying open questions around specification, auditing, and deployment.
- Certifications
- AI policy
- Quality assurance
Research
ARTIFICIAL INTELLIGENCE ADOPTION AND ORGANIZATIONAL PERFORMANCE AMONG SMALL AND MEDIUM ENTERPRISES IN UYO METROPOLIS, AKWA IBOM STATE, NIGERIA
Gaius-Okeh Happiness Adanneya PhD, Ogechi Anastasia Emenike, Ifelunwa Ada Ikuni
Zenodo (CERN European Organization for Nuclear Research) · 2026-06-22
This study of 312 SME owner-managers in Uyo, Nigeria found a strong positive relationship between AI adoption and organizational performance (r = 0.68), with operational, analytical, and generative AI tools jointly explaining 54% of the variance in performance outcomes. Operational AI had the strongest individual effect (β = 0.41), suggesting that automation of day-to-day business processes yields the greatest performance gains for small firms. The findings highlight that AI adoption is a significant driver of SME growth in sub-national Nigerian markets, where empirical evidence has previously been scarce. The authors recommend targeted digital-skills training, infrastructure improvements, and government-backed financing to support responsible AI adoption among small businesses.
- Enterprise
- Workforce
- AI policy
Research
The Urban-Rural Divide in the Age of Artificial Intelligence: Assessing the Effects of Technology and Automation on Regional Labor Markets
CHAU TRAN BAO, Khoi Nguyen Dinh Nguyen, Ha Nguyen Manh et al.
arXiv (Cornell University) · 2026-06-22
This study examines how automation and AI differentially affect employment and wages across urban and rural labor markets using panel data with two-way fixed-effects and instrumental-variable models. The findings show that automation exposure—concentrated in routine work—lowers employment and wages, with the employment losses somewhat cushioned in cities, while AI exposure—concentrated in cognitive work—raises wages but is concentrated in urban areas. The research argues that technology reshapes rather than simply widens the urban-rural divide, and calls for place-sensitive workforce and education policy that targets reskilling support toward routine-exposed rural regions and extends digital infrastructure outward so rural workers can benefit from AI's wage gains.
- Workforce
- AI policy
Research
Are Safety Guarantees in Neural Networks Safe? How to Compute Trustworthy Robustness Certifications
Merkouris Papamichail, Konstantinos Varsos, Giorgos Flouris et al.
arXiv (Cornell University) · 2026-06-22
This paper addresses a fundamental AI safety challenge: verifying that neural networks are robust to adversarial examples—slightly distorted inputs that cause misclassification. The authors introduce the 'apothem measure' for computing robustness certifications and prove that volume-optimal certifications are computationally intractable even with idealized oracle access, while their approach achieves apothem-optimal certifications in a linear number of oracle calls. Their system, ParallelepipedoNN, is evaluated on MNIST and Fashion MNIST benchmarks, achieving at least a two-fold improvement over existing methods in minimum edge length of the certified region. These findings matter because trustworthy robustness guarantees are essential for certifying that AI systems behave safely under real-world input perturbations.
- Quality assurance
- Certifications
Research
Reliability Claims in AI-Driven Assistive Technology, and the Case for Independent Certification
Christopher Hamilton
Zenodo (CERN European Organization for Nuclear Research) · 2026-06-22
This paper examines the risks posed by unverified reliability claims in AI-driven assistive technologies, where users with disabilities may be unable to independently check system outputs—making errors both safety-critical and silent. It introduces a taxonomy of claim types (asserted, scoped, and verified) and draws on the 2025 FTC action against an accessibility-overlay vendor to illustrate real-world harms. The authors argue that responsible practice requires not eliminating errors but governing them through scoped claims, output verification, mandatory human review, and independent certification against published standards. Practical recommendations are offered for vendors, buyers, and standards bodies.
- Certifications
- Quality assurance
- AI policy
Research
Ten Digits on a Train: AI-Assisted Verification of Two Eigenvalue Problems
Matthew J. Colbrook
arXiv (Cornell University) · 2026-06-22
This paper reports a human-AI collaboration to certify eigenvalue computations to ten decimal places in two mathematically challenging settings: a singular self-adjoint Schrödinger operator and a non-normal atom-molecule resonance benchmark. The authors develop a reusable verified-computation architecture using a Krawczyk-Brouwer inclusion method that handles ill-conditioned propagation and uncertain asymptotic data. The work demonstrates both the speed and the limits of AI assistance—AI rapidly generated candidates and proof strategies, but human judgment was required to catch critical errors in AI-proposed arguments. The authors argue that as AI lowers the cost of code and numerical claims, standards for verification, attribution, peer review, and training data must adapt accordingly.
- Quality assurance
- Certifications
- AI policy
Research
Reliability Claims in AI-Driven Assistive Technology, and the Case for Independent Certification
Christopher Hamilton
Zenodo (CERN European Organization for Nuclear Research) · 2026-06-22
This paper examines the risks posed by unverifiable reliability claims in AI-driven assistive technologies, where users with disabilities may be unable to independently check system outputs. It introduces a taxonomy of claim types—asserted, scoped, and verified—and argues that blanket reliability assertions are especially dangerous in assistive contexts where errors can be safety-critical and silent. Drawing on the 2025 FTC action against an accessibility-overlay vendor and the architecture of hybrid AI systems, the authors contend that the right response is not eliminating error but governing it through scoped claims, output verification, mandatory human review, periodic re-testing, and independent certification against published standards. The paper closes with practical recommendations for vendors, buyers, and standards bodies.
- Certifications
- AI policy
- Quality assurance
Research
ARTIFICIAL INTELLIGENCE ADOPTION AND ORGANIZATIONAL PERFORMANCE AMONG SMALL AND MEDIUM ENTERPRISES IN UYO METROPOLIS, AKWA IBOM STATE, NIGERIA
Gaius-Okeh Happiness Adanneya PhD, Ogechi Anastasia Emenike, Ifelunwa Ada Ikuni
Zenodo (CERN European Organization for Nuclear Research) · 2026-06-22
This study surveyed 312 SME owner-managers in Uyo, Nigeria to assess whether AI adoption drives measurable organizational performance gains. Using Pearson correlation and multiple regression, the researchers found a strong positive relationship between AI adoption and performance (r = 0.68), with operational, analytical, and generative AI together explaining 54% of the variance in performance (R² = 0.54). Operational AI had the strongest individual effect (β = 0.41). The authors recommend digital-skills training, infrastructure investment, and government financing schemes to accelerate responsible AI adoption among small and medium enterprises.
- Enterprise
- Workforce
- AI policy