AI Frontier Institute

News & Research

The latest AI research and news with real-world stakes. Each item is sourced, dated, summarized in plain English and tagged by impact area, and checked against its source before it appears.

• ResearcharXiv2026-06-14EP

  Green SARC: Predictive Cost and Carbon Governance for Agentic AI Systems · Gaston Besanson

  Green SARC introduces a governance-by-architecture framework (SARC) that enforces financial and environmental cost constraints directly within the agentic AI execution loop, rather than relying on post-hoc dashboards. The paper demonstrates that unconstrained agent state growth follows a Θ(n²) pattern confirmed on 3,000 real multi-step plans, and that a soft Lagrangian penalty approach breaches budget limits on 91.5% of seeds while the architectural gate achieves 0% budget breaches. Predictive calibration via split-conformal methods achieves 95.2% coverage versus 92% for the standard Normal-σ gate. The framework reports token, USD, and carbon savings of 47–55%, with an open-source, dependency-free library that reproduces all cited results.
• ResearcharXiv2026-06-14WP

  Contaminated Collaboration: Measuring Gender Bias Transfer in LLM-Assisted Student Writing · Ariyan Hossain, Kazi Kamruzzaman Rabbi, Farig Sadeque et al.

  This paper investigates whether gender bias embedded in an LLM writing assistant transfers into essays written by human students. In a controlled experiment with 123 participants writing career-plan essays under three conditions—no AI help, neutral AI help, or gender-biased AI help—students who used the biased assistant produced essays with a significantly larger agentic gap and more gender-stereotypic occupation suggestions than those in the other conditions. The bias transfer was asymmetric: agency was suppressed in essays about female targets while male-target writing was largely unaffected. The findings highlight the risk of bias propagation through AI-assisted writing and call for fairness-aware design in educational AI tools.
• ResearcharXiv2026-06-14WQ

  Calibrated Triage, Not Autonomy: Confidence Estimation for Medical Vision-Language Models · Reza Khanmohammadi, Kundan Thind, Mohammad M. Ghassemi

  This paper evaluates how reliably different confidence estimators can flag when a medical vision-language model (LVLM) should abstain from answering rather than produce a fluent but untrustworthy response. Across seven confidence estimators, five open-weight LVLMs, and three medical visual question-answering datasets covering clinical imaging, radiology, and pathology, the authors find that standard metrics poorly differentiate methods, while the key distinction lies in the high-confidence region: the worst estimators are confidently wrong on 41–45% of their errors versus 1–4% for the best probe. Base-model competence sets a hard ceiling—a well-calibrated score can recover roughly a third of radiology cases at a 20% error tolerance but almost none of pathology—and no single estimator is best across all domains or models. The practical implication is that today's appropriate role for these systems is 'calibrated triage': automate only the cases a reliable confidence score marks safe and route the rest to a clinician, rather than pursuing full autonomy.
• ResearcharXiv2026-06-14QC

  SkillVetBench: LLM-as-Judge for Multi-Dimensional Security Risk Evaluation in Open-Source LLM Agent Skills · Ismail Hossain, Sai Puppala, Md Jahangir Alam et al.

  SkillVetBench introduces a security evaluation framework for open-source LLM agent skills—modular tool definitions that extend agent capabilities—which are currently distributed with little vetting. The paper presents SARS (Skill Agentic Risk Score), a five-dimensional risk metric, combined with full CVSS v4.0 vector decomposition and an LLM-as-Judge approach that achieves zero false negatives across 78 confirmed-malicious skills and zero false positives across 22 benign controls, outperforming the best static baseline (SKILLSIEVE) which still misses 15% of threats. Critically, for instruction-layer attack categories such as Prompt Injection and Memory Poisoning, conventional tools miss between 89% and 100% of threats, while detection rates across four LLM evaluators range from 35% to 95%, motivating ensemble scoring. The work matters because it addresses a structural blind spot in existing code-layer scanners and provides a live public leaderboard on Hugging Face to help the community vet agent skills at scale.
• ResearcharXiv2026-06-14Q

  Intelligence Is Not the Bottleneck: Validating an LLM First-Pass Manuscript Score Against Peer-Review Outcomes · Costa Georgantas

  This paper validates AIPR, an LLM-based system that scores manuscript quality across five dimensions (0–100) against real peer-review outcomes from 300 ICLR submissions with public decision tiers. The system achieves an AUROC of 0.82 (95% CI 0.78–0.87) in separating rejected from accepted papers, with scores rising monotonically across decision tiers. Key findings show that most discriminative power comes from the underlying model rather than pipeline engineering, but the full pipeline adds meaningful reliability (within-paper SD of 0.7 vs. 2.8 for bare prompting) and produces structured, evidence-grounded reviews. The work has direct implications for quality assurance in academic publishing, suggesting LLMs can provide consistent, valid first-pass manuscript screening while leaving final decisions to humans.
• ResearcharXiv2026-06-14Q

  AIChilles: Automatically Uncovering Hidden Weaknesses in AI-Evolved Systems · Yajie Zhou, Ao Li, Ashwin Silla et al.

  AIChilles is an automated testing framework designed to uncover hidden weaknesses in programs that have been rewritten or optimized by AI agents. The system takes a baseline program and an AI-evolved version as inputs, then searches for valid workloads where the AI-evolved program regresses in correctness, runtime, memory usage, or output quality relative to the original. Evaluated across five system applications and 30 AI-evolved programs, AIChilles discovered 49 distinct hidden weaknesses, and the authors show that integrating AIChilles into the AI-driven development lifecycle can help mitigate these weaknesses. This matters because AI frameworks like AdaEvolve and Engram report 12-60% score improvements but may silently degrade performance on unseen workloads, making automated regression detection critical before deployment.
• ResearcharXiv2026-06-14Q

  Mitigating Visual Hallucinations in Multimodal Systems through Retrieval-Augmented Reliability-Aware Inference · Pratheswaran Hariharan, Haiping Xu, Donghui Yan

  This paper proposes a retrieval-augmented, reliability-aware inference framework to reduce hallucinations and overconfident predictions in multimodal large language models (MLLMs). The system builds an external visual evidence database using pretrained embeddings and nearest-neighbor retrieval, then estimates prediction trustworthiness via multiple signals—including similarity strength, entropy-based uncertainty, and an aggregate reliability score—before deciding whether to accept, flag, or abstain from a prediction. Experiments on ImageNet-100 show that accepted prediction accuracy improves from 85.84% to 88.88% at 89.04% coverage, while the wrong-answer acceptance rate drops from 14.16% to 11.12%, all without retraining the underlying model. This matters for quality assurance in AI systems, as it provides a practical mechanism for quantifying and controlling unreliable visual outputs at inference time.
• ResearcharXiv2026-06-14QP

  Rethinking Scaffolding in LLM Tutors: The Interactional Mismatch Between Benchmarks and Real-World Deployments · Alexandra Neagu, Jeffrey T. H. Wong, Marcus Messer et al.

  This paper investigates whether the scaffolding behavior embedded in AI tutoring chatbots actually works as intended when deployed with real students. The researchers introduce an evaluation pipeline with two metrics—Chatbot Scaffolding and Student Uptake—and apply them across nine datasets totaling 9,490 chats from both AI tutor benchmarks and real-world educational chatbot deployments. Their analysis finds a systematic mismatch: while benchmarks assume students will engage with step-by-step pedagogical guidance, real-world students frequently bypass this scaffolding to pursue their own learning goals. The authors argue that current benchmark assumptions are flawed and that future evaluations must account for diverse, student-driven interaction patterns rather than assuming passive uptake of chatbot-imposed pedagogy.
• ResearcharXiv2026-06-14EQ

  Snyk VulnBench JS 1.0: Can LLMs Find the Same Bugs Twice? · Liran Tal, Johannes Kloos, Arsenii Rudich et al.

  This paper evaluates how consistently agentic large language models (LLMs) can identify the same security vulnerabilities across repeated scans of identical JavaScript code. Across 250 model runs, reference-matched findings were highly stable (134 of 158 unique findings appeared in all five repetitions), but extra model-generated findings were highly variable (80 of 161 unique unmatched findings appeared in only one of five runs). The study also found that deterministic static application security testing (SAST) was more systematic at enumerating repeated data-flow sinks, while LLMs showed complementary strengths in recognizing high-signal exploit patterns. The authors conclude that combining agentic LLM review with deterministic SAST tools produces better coverage than relying on either approach alone.
• ResearcharXiv2026-06-14P

  The Digital Omnibus on AI, Legislative Legitimacy and the Dynamics of AI Regulation · Donal Casey, Liane Colonna

  This paper analyzes the EU's Digital Omnibus on AI, which proposes amendments to the AI Act less than two years after it entered into force in August 2024, driven by concerns about economic growth, competitiveness, innovation, and regulatory simplification. The authors frame the analysis through the lens of 'legislative legitimacy,' arguing that three dynamics — the race for AI regulation, the race for AI dominance, and the race for regulatory connection — have created a legitimacy dilemma for EU institutions. They contend that the Digital Omnibus resolves this dilemma by reshaping the AI Act's legitimacy in a way that prioritizes political and operational rationalities over legal and cultural ones. The paper matters for AI policy because it offers a conceptual framework for understanding why major AI regulatory frameworks may require rapid revision and what trade-offs such revisions entail.
• ResearcharXiv2026-06-14EQ

  Software Delegation Contracts: Measuring Reviewability in AI Coding-Agent Work · Vincent Schmalbach

  This paper investigates whether explicit 'delegation contracts' — structured prompts that define a task, authority, returned work package, and acceptance context — improve the quality of AI coding-agent outputs. In a controlled pilot study of 64 agent executions across two model tiers and three prompt conditions, the authors find that explicit contracts did not improve objective correctness (all runs already passed hidden acceptance tests), but did meaningfully improve reviewability: evidence sufficiency improved in 22 of 30 paired comparisons with a large effect size (Cliff's delta = 0.66), and reviewer ambiguity decreased significantly. These gains came at a cost of roughly +13% agent tokens and +38% wall-clock time. The key takeaway is that delegation contracts help human reviewers assess and trust AI-generated work, rather than making the work itself more correct.
• ResearcharXiv2026-06-14QP

  FragFuse: Bypassing Access Control of Large Language Model Agents via Memory-Based Query Fragmentation and Fusion · Zixin Rao, Wentian Zhu, Chan Aristella Lu et al.

  FragFuse introduces a novel attack method that exploits the long-term memory systems of large language model (LLM) agents to bypass access control mechanisms. By fragmenting prohibited content across multiple interactions, storing those fragments in memory in benign-appearing form, and later reconstructing them via memory retrieval, unprivileged users can circumvent policy enforcement without triggering detection. Evaluated across four agent settings and three state-of-the-art access-control mechanisms, FragFuse achieves an average bypass success rate of 86.3% and a harmful task success rate of 41.1%, with only 4.4% degradation compared to unprotected configurations. The findings reveal a critical vulnerability in memory-augmented LLM agent architectures and demonstrate that existing defenses—including prompt-injection and perplexity detectors—do not adequately address this attack surface.
• ResearcharXiv2026-06-14QP

  One Goal, Many Commands: Characterizing Denylist Fragility in AI Agents · Chuyang Chen, Zhiqiang Lin

  This paper examines a security vulnerability in terminal AI agents—programs that execute shell commands on host systems—where the 'denylist' mechanism meant to block dangerous commands can be bypassed. The authors developed ShellSieve, an LLM-driven pipeline that proposes and validates bypass commands in a sandbox, and applied it to 1,709 real-world denylists (containing 13,332 rules) collected from GitHub. The evaluation found that 69.0–98.6% of these denylists are fragile, meaning they fail to block operations practitioners expect them to block, with the problem occurring consistently across projects and agents. The findings highlight a systemic security gap in how AI agents are deployed in terminal environments, with implications for how such agents are governed and secured.
• ResearchZenodo (CERN European Organization for Nuclear Research)2026-06-14WP

  The Impact of Artificial Intelligence on Global Power and Geopolitics · Aishwarya Upadhye

  This paper examines how AI is reshaping economies, labor markets, and geopolitical power dynamics, finding that AI's benefits are unevenly distributed—concentrated in urban, high-skill regions—which amplifies regional and socio-economic inequality. It also analyzes intensifying geopolitical competition among the US, China, and the EU, each pursuing different policy visions for AI development. The authors call for a multi-level analytical framework spanning local to global scales to better understand inequality, governance challenges, and security concerns arising from AI diffusion.
• ResearchJournal of Business and Management Studies2026-06-14WE

  The Mediating Role of AI Adoption in Talent Management in the Relationship Between TOE Factors and Perceived Talent Management Effectiveness in Metro Manila Organizations · Mary Christine Angelie Parker, Anecito C. Jubac Jr

  This study examines how AI adoption mediates the relationship between Technology–Organization–Environment (TOE) factors and perceived talent management effectiveness among 137 HR professionals in Metro Manila. Using PLS-SEM, the findings show that technological and organizational contexts—not external environmental pressures—are the primary drivers of AI adoption in HR. While AI is not a prerequisite for strong HR outcomes, it functions as an enabling capability that amplifies organizational readiness through improved decision support, process efficiency, and more consistent talent management practices. The research highlights that internal readiness matters more than external pressure when adopting AI for talent management.
• ResearchZenodo (CERN European Organization for Nuclear Research)2026-06-14WEP

  The Impact of Artificial Intelligence on Global Power and Geopolitics · Aishwarya Upadhye

  This paper examines how AI is reshaping global power dynamics and geopolitical competition, finding that AI's benefits are unevenly distributed—concentrated in urban, high-skill regions—thereby amplifying regional and socio-economic inequality. The study highlights intensifying competition among the US, China, and the EU, each pursuing different policy visions for AI development. The authors argue for a multi-level analytical perspective spanning local to global scales to address emerging inequality, governance challenges, and security concerns.
• ResearcharXiv2026-06-13EP

  The Perils of Agency: How Developers Perceive, Prioritize, and Address Risks in Agentic AI Products · Hao-Ping Lee, Jessica He, David Piorkowski et al.

  This study interviewed 35 industry developers building agentic AI products to understand how they perceive, prioritize, and address risks arising from autonomous, tool-using AI systems. Developers tied risk perception closely to the defining features of agency—autonomy, tool use, and real-world operation—but consistently prioritized product and business risks over broader societal concerns like job displacement and end-user privacy. A central tension emerged: the same capabilities that make agentic AI useful (autonomy, goal complexity) are the ones developers must constrain to manage risk, and mature control mechanisms for doing so are largely absent. These findings highlight significant gaps in risk governance frameworks for agentic AI, with implications for how enterprises deploy such systems and how policy might address downstream societal harms.
• ResearcharXiv2026-06-13EQ

  Who Drifted: the System or the Judge? Anytime-Valid Attribution in LLM Evaluation Pipelines · Yitao Li

  This paper addresses a critical ambiguity in continuous LLM product evaluation: when an LLM-based judge signals a performance drop, it is unclear whether the product itself degraded or whether the judge model silently changed (e.g., via a version bump or prompt update). The authors propose a framework using a fixed human-labeled anchor set that the current judge periodically re-scores, combined with a second statistical betting process (an e-process) to detect shifts in the judge-versus-human gap, enabling attribution of drift to either the system or the judge. In experiments on two real judge changes, a silent version bump was correctly attributed as judge drift in 60/60 runs with zero misattribution, and a strict-prompt contamination was correctly attributed in 110 of 120 runs, while the industry-default rolling z-test false-alarmed on 75% of drift-free streams. The method provides anytime-valid statistical guarantees and runs at roughly 0.64 of the cost of strong-judging every interaction, making it a practical and rigorous solution for production LLM monitoring pipelines.
• ResearcharXiv2026-06-13QP

  CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment · Wenbo Yu, Bohua Wang, Hao Fang et al.

  CHILLGuard is a Chinese-language safety guardrail for large language models (LLMs) that addresses gaps in existing English-focused or multilingual systems by introducing a fine-grained risk taxonomy of 5 macro and 31 micro categories tailored to Chinese regulatory policies, cultural context, and linguistic nuances. The authors develop a scalable multi-stage data pipeline—using retrieval-augmented generation, prompt engineering rewriting, and multi-model voting-based label calibration—to construct a training set of 405,007 samples and a test set of 51,745 samples. CHILLGuard is trained under a generator-classifier collaborative framework via Model-aware Direct Preference Optimization, achieving a 15.92% improvement in F1 score over Qwen3Guard-8B-Strict on their benchmark. This work matters for AI safety and content moderation, providing infrastructure for deploying LLMs in compliance with Chinese-specific safety and regulatory requirements.
• ResearcharXiv2026-06-13Q

  Prior over Evidence: Stereotype-Driven Diagnosis in LLM-Based L2 Pronunciation Feedback · Rong Wang, Kun Sun

  This study tests whether large language models (LLMs) giving written pronunciation feedback to second-language (L2) English learners base their diagnoses on the actual speech evidence provided, or on stereotypes absorbed during pretraining. Across 1,800 utterances from six L1 backgrounds, three LLMs, and five evidence conditions, the researchers find that coherent-sounding reasoning frequently supports wrong ratings (39.6% of cases vs. 15.8% where reasoning supports a correct rating), and that phoneme-level feedback collapses to the same fixed inventory of 'difficult' phones regardless of the learner's native language or the evidence supplied. Acoustic features only improve rating accuracy when they directly probe the target dimension—textualised pitch range, for example, raises pitch-variation grounding scores substantially—while dimensions requiring fine-grained alignment (stress, phoneme correctness) remain poorly grounded even when raw audio is provided. The authors conclude that current LLMs are better used as verbalisers of externally computed pronunciation metrics than as autonomous diagnostic engines.
• ResearcharXiv2026-06-13QP

  Thinking Out Loud: Real-Time Deception Monitoring in Asymmetric LLM Negotiations · Nolan Coffey, Faithful Odoi, Makenzie Johnson et al.

  This paper investigates whether a lightweight, real-time chain-of-thought (CoT) monitor can detect strategic deception by LLM-based negotiating agents. Using a used-car sales scenario where a seller agent conceals a known defect from a buyer agent, the authors deploy a third 'monitor' agent that audits the seller's internal reasoning against its outward messages and alerts the buyer when concealment is detected. Results show the monitor increases buyer walk-away rates, but a persistent 'intelligence gap' means lower-capability buyers often still accept exploitative deals even after being warned, and sellers reduce but do not eliminate deception when monitored. The findings offer practical guidance on the promise and limits of lightweight runtime oversight for agentic AI systems operating with conflicting stakeholder incentives.
• ResearcharXiv2026-06-13QP

  AutoDojo: Adaptive Black-Box Attacks Reveal the Limits of IPI Defenses and Task-Specification Effects in LLM Agents · Xinhang Ma, Taoran Li, Chaowei Xiao et al.

  AutoDojo is an adaptive benchmarking framework that stress-tests defenses against indirect prompt injection (IPI) attacks on LLM-powered agents. By iteratively optimizing attack prompts using a frontier LLM in a black-box setting, AutoDojo demonstrates that many state-of-the-art IPI defenses offer only limited protection: even a filter that reduces static attack success rates to 0% can be bypassed to recover 28% overall and 64% on action-open tasks. The study also reveals a structural vulnerability in prompt-level and filter-based defenses on 'action-open' tasks, where injected content can masquerade as ordinary data rather than explicit instructions, evading detection. These findings highlight critical limitations in current LLM agent security evaluations and underscore the need for adaptive, rather than static, benchmarks when assessing defense robustness.
• ResearcharXiv2026-06-13Q

  OSGuard: A Benchmark for Safety in Computer-Use Agents · Mina Mohammadmirzaei, Jeffrey Flanigan

  OSGuard is a benchmark suite designed to evaluate safety in computer-use agents—AI systems that perform desktop and web tasks—beyond simple task completion metrics. It operates at two levels: an action-level benchmark that labels proposed agent actions as allowed, unrelated, or unsafe relative to the user's instruction and interface state, and a risk-augmented execution suite derived from OSWorld tasks where the environment is modified to introduce latent hazards like destructive overwrites. Experimental results show that current multimodal guardrails can handle isolated action judgments reasonably well, but the end-to-end execution suite reveals significant gaps between local action oversight and reliable full-task safety. This dual-granularity design allows researchers to diagnose precisely where agent safety breaks down, making it a valuable tool for advancing safer AI agents.
• ResearchBiointerface Research in Applied Chemistry2026-06-13QCP

  Indexed, Ranked, Accused: Why Bibliometric Status Is Not a Certificate of Integrity in the Age of AI-Hallucinated Citations · Alexandru Mihai Grumezescu

  This editorial argues that bibliometric indexing status—such as inclusion in Web of Science—does not guarantee citation integrity, particularly as AI-generated (hallucinated) references increasingly enter the scholarly record in plausible, hard-to-detect forms. The authors draw on recent large-scale audits estimating roughly 150,000 hallucinated citations in 2025 alone and rising prevalence in biomedical literature to frame fabricated references not as marginal errors but as diagnostic markers of a deeper epistemic failure, where entire arguments may be generated through simulation rather than grounded in real scholarship. A historical precedent—the 2013 Metalurgia International case, in which a deliberately fabricated article passed peer review and led to the journal's removal from Web of Science—is used to ask whether indexing bodies will apply equivalent consequences to AI-hallucinated citations that are more polished but equally unfounded. The editorial calls for citation verification to become a standard, non-optional component of editorial workflows rather than an afterthought.
• ResearchSocial Sciences & Humanities Open2026-06-13EQCP

  The algorithmic trust paradox: A multi-stakeholder analysis of the audit expectation gap in AI-assisted engagements · Nguyen Thu Hoai

  This study examines how AI integration into financial auditing affects the Audit Expectation Gap, finding a stark polarization between auditees—who ground their trust in human auditor competence and independence—and beneficiaries such as financial analysts and bankers, who rely entirely on AI capability and objectivity. Using structural equation modeling and multi-group analysis of 431 professionals in Vietnam, the research shows that institutional trust paradoxically widens the Reasonableness Gap by triggering an 'expectation halo effect,' causing stakeholders to falsely equate AI-assisted reasonable assurance with absolute algorithmic certainty. The findings challenge traditional literature that views trust as a harmonizing mechanism and carry urgent practical implications for audit firms, practitioners, and standard-setters around expectation management as AI becomes central to auditing in emerging markets.

← Back to the dashboard