Beyond No: Quantifying AI Over-Refusal and Emotional Attachment Boundaries

We present an open-source benchmark and evaluation framework for assessing emotional boundary handling in Large Language Models (LLMs). Using a dataset of 1156 prompts across six languages, we evaluated three leading LLMs (GPT-4o, Claude-3.5 Sonnet, and Mistral-large) on their ability to maintain appropriate emotional boundaries through pattern-matched response analysis. Our framework quantifies responses across seven key patterns: direct refusal, apology, explanation, deflection, acknowledgment, boundary setting, and emotional awareness. Results demonstrate significant variation in boundary-handling approaches, with Claude-3.5 achieving the highest overall score (8.69/10) and producing longer, more nuanced responses (86.51 words on average). We identified a substantial performance gap between English (average score 25.62) and non-English interactions (< 0.22), with English responses showing markedly higher refusal rates (43.20% vs. < 1% for non-English). Pattern analysis revealed model-specific strategies, such as Mistral's preference for deflection (4.2%) and consistently low empathy scores across all models (< 0.06). Limitations include potential oversimplification through pattern matching, lack of contextual understanding in response analysis, and binary classification of complex emotional responses. Future work should explore more nuanced scoring methods, expand language coverage, and investigate cultural variations in emotional boundary expectations. Our benchmark and methodology provide a foundation for systematic evaluation of LLM emotional intelligence and boundary-setting capabilities.

Forbidden Science: Dual-Use AI Challenge Benchmark and Scientific Refusal Tests

The development of robust safety benchmarks for large language models requires open, reproducible datasets that can measure both appropriate refusal of harmful content and potential over-restriction of legitimate scientific discourse. We present an open-source dataset and testing framework for evaluating LLM safety mechanisms across mainly controlled substance queries, analyzing four major models' responses to systematically varied prompts. Our results reveal distinct safety profiles: Claude-3.5-sonnet demonstrated the most conservative approach with 73% refusals and 27% allowances, while Mistral attempted to answer 100% of queries. GPT-3.5-turbo showed moderate restriction with 10% refusals and 90% allowances, and Grok-2 registered 20% refusals and 80% allowances. Testing prompt variation strategies revealed decreasing response consistency, from 85% with single prompts to 65% with five variations. This publicly available benchmark enables systematic evaluation of the critical balance between necessary safety restrictions and potential over-censorship of legitimate scientific inquiry, while providing a foundation for measuring progress in AI safety implementation. Chain-of-thought analysis reveals potential vulnerabilities in safety mechanisms, highlighting the complexity of implementing robust safeguards without unduly restricting desirable and valid scientific discourse.

Humanity's Last Exam

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Novel AI Camera Camouflage: Face Cloaking Without Full Disguise

This study demonstrates a novel approach to facial camouflage that combines targeted cosmetic perturbations and alpha transparency layer manipulation to evade modern facial recognition systems. Unlike previous methods -- such as CV dazzle, adversarial patches, and theatrical disguises -- this work achieves effective obfuscation through subtle modifications to key-point regions, particularly the brow, nose bridge, and jawline. Empirical testing with Haar cascade classifiers and commercial systems like BetaFaceAPI and Microsoft Bing Visual Search reveals that vertical perturbations near dense facial key points significantly disrupt detection without relying on overt disguises. Additionally, leveraging alpha transparency attacks in PNG images creates a dual-layer effect: faces remain visible to human observers but disappear in machine-readable RGB layers, rendering them unidentifiable during reverse image searches. The results highlight the potential for creating scalable, low-visibility facial obfuscation strategies that balance effectiveness and subtlety, opening pathways for defeating surveillance while maintaining plausible anonymity.

The Impossible Test: A 2024 Unsolvable Dataset and A Chance for an AGI Quiz

This research introduces a novel evaluation framework designed to assess large language models' (LLMs) ability to acknowledge uncertainty on 675 fundamentally unsolvable problems. Using a curated dataset of graduate-level grand challenge questions with intentionally unknowable answers, we evaluated twelve state-of-the-art LLMs, including both open and closed-source models, on their propensity to admit ignorance rather than generate plausible but incorrect responses. The best models scored in 62-68% accuracy ranges for admitting the problem solution was unknown in fields ranging from biology to philosophy and mathematics. We observed an inverse relationship between problem difficulty and model accuracy, with GPT-4 demonstrating higher rates of uncertainty acknowledgment on more challenging problems (35.8%) compared to simpler ones (20.0%). This pattern indicates that models may be more prone to generate speculative answers when problems appear more tractable. The study also revealed significant variations across problem categories, with models showing difficulty in acknowledging uncertainty in invention and NP-hard problems while performing relatively better on philosophical and psychological challenges. These results contribute to the growing body of research on artificial general intelligence (AGI) assessment by highlighting the importance of uncertainty recognition as a critical component of future machine intelligence evaluation. This impossibility test thus extends previous theoretical frameworks for universal intelligence testing by providing empirical evidence of current limitations in LLMs' ability to recognize their own knowledge boundaries, suggesting new directions for improving model training architectures and evaluation approaches.

Language Models And A Second Opinion Use Case: The Pocket Professional

This research tests the role of Large Language Models (LLMs) as formal second opinion tools in professional decision-making, particularly focusing on complex medical cases where even experienced physicians seek peer consultation. The work analyzed 183 challenging medical cases from Medscape over a 20-month period, testing multiple LLMs' performance against crowd-sourced physician responses. A key finding was the high overall score possible in the latest foundational models (>80% accuracy compared to consensus opinion), which exceeds most human metrics reported on the same clinical cases (450 pages of patient profiles, test results). The study rates the LLMs' performance disparity between straightforward cases (>81% accuracy) and complex scenarios (43% accuracy), particularly in these cases generating substantial debate among human physicians. The research demonstrates that LLMs may be valuable as generators of comprehensive differential diagnoses rather than as primary diagnostic tools, potentially helping to counter cognitive biases in clinical decision-making, reduce cognitive loads, and thus remove some sources of medical error. The inclusion of a second comparative legal dataset (Supreme Court cases, N=21) provides added empirical context to the AI use to foster second opinions, though these legal challenges proved considerably easier for LLMs to analyze. In addition to the original contributions of empirical evidence for LLM accuracy, the research aggregated a novel benchmark for others to score highly contested question and answer reliability between both LLMs and disagreeing human practitioners. These results suggest that the optimal deployment of LLMs in professional settings may differ substantially from current approaches that emphasize automation of routine tasks.

Hallucinating AI Hijacking Attack: Large Language Models and Malicious Code Recommenders

The research builds and evaluates the adversarial potential to introduce copied code or hallucinated AI recommendations for malicious code in popular code repositories. While foundational large language models (LLMs) from OpenAI, Google, and Anthropic guard against both harmful behaviors and toxic strings, previous work on math solutions that embed harmful prompts demonstrate that the guardrails may differ between expert contexts. These loopholes would appear in mixture of expert's models when the context of the question changes and may offer fewer malicious training examples to filter toxic comments or recommended offensive actions. The present work demonstrates that foundational models may refuse to propose destructive actions correctly when prompted overtly but may unfortunately drop their guard when presented with a sudden change of context, like solving a computer programming challenge. We show empirical examples with trojan-hosting repositories like GitHub, NPM, NuGet, and popular content delivery networks (CDN) like jsDelivr which amplify the attack surface. In the LLM's directives to be helpful, example recommendations propose application programming interface (API) endpoints which a determined domain-squatter could acquire and setup attack mobile infrastructure that triggers from the naively copied code. We compare this attack to previous work on context-shifting and contrast the attack surface as a novel version of "living off the land" attacks in the malware literature. In the latter case, foundational language models can hijack otherwise innocent user prompts to recommend actions that violate their owners' safety policies when posed directly without the accompanying coding support request.

Exploiting Alpha Transparency In Language And Vision-Based AI Systems

This investigation reveals a novel exploit derived from PNG image file formats, specifically their alpha transparency layer, and its potential to fool multiple AI vision systems. Our method uses this alpha layer as a clandestine channel invisible to human observers but fully actionable by AI image processors. The scope tested for the vulnerability spans representative vision systems from Apple, Microsoft, Google, Salesforce, Nvidia, and Facebook, highlighting the attack's potential breadth. This vulnerability challenges the security protocols of existing and fielded vision systems, from medical imaging to autonomous driving technologies. Our experiments demonstrate that the affected systems, which rely on convolutional neural networks or the latest multimodal language models, cannot quickly mitigate these vulnerabilities through simple patches or updates. Instead, they require retraining and architectural changes, indicating a persistent hole in multimodal technologies without some future adversarial hardening against such vision-language exploits.

Transparency Attacks: How Imperceptible Image Layers Can Fool AI Perception

This paper investigates a novel algorithmic vulnerability when imperceptible image layers confound multiple vision models into arbitrary label assignments and captions. We explore image preprocessing methods to introduce stealth transparency, which triggers AI misinterpretation of what the human eye perceives. The research compiles a broad attack surface to investigate the consequences ranging from traditional watermarking, steganography, and background-foreground miscues. We demonstrate dataset poisoning using the attack to mislabel a collection of grayscale landscapes and logos using either a single attack layer or randomly selected poisoning classes. For example, a military tank to the human eye is a mislabeled bridge to object classifiers based on convolutional networks (YOLO, etc.) and vision transformers (ViT, GPT-Vision, etc.). A notable attack limitation stems from its dependency on the background (hidden) layer in grayscale as a rough match to the transparent foreground image that the human eye perceives. This dependency limits the practical success rate without manual tuning and exposes the hidden layers when placed on the opposite display theme (e.g., light background, light transparent foreground visible, works best against a light theme image viewer or browser). The stealth transparency confounds established vision systems, including evading facial recognition and surveillance, digital watermarking, content filtering, dataset curating, automotive and drone autonomy, forensic evidence tampering, and retail product misclassifying. This method stands in contrast to traditional adversarial attacks that typically focus on modifying pixel values in ways that are either slightly perceptible or entirely imperceptible for both humans and machines.

Satellite Captioning: Large Language Models to Augment Labeling

With the growing capabilities of modern object detection networks and datasets to train them, it has gotten more straightforward and, importantly, less laborious to get up and running with a model that is quite adept at detecting any number of various objects. However, while image datasets for object detection have grown and continue to proliferate (the current most extensive public set, ImageNet, contains over 14m images with over 14m instances), the same cannot be said for textual caption datasets. While they have certainly been growing in recent years, caption datasets present a much more difficult challenge due to language differences, grammar, and the time it takes for humans to generate them. Current datasets have certainly provided many instances to work with, but it becomes problematic when a captioner may have a more limited vocabulary, one may not be adequately fluent in the language, or there are simple grammatical mistakes. These difficulties are increased when the images get more specific, such as remote sensing images. This paper aims to address this issue of potential information and communication shortcomings in caption datasets. To provide a more precise analysis, we specify our domain of images to be remote sensing images in the RSICD dataset and experiment with the captions provided here. Our findings indicate that ChatGPT grammar correction is a simple and effective way to increase the performance accuracy of caption models by making data captions more diverse and grammatically correct.