Import AI 479: Spacewar; low-latency tasks versus LLMs; Genie 4
What do LLMs think about vision models?
What do LLMs think about vision models?
A growing number of startups are editing powerful Chinese open-weight models for Western businesses that are concerned about censorship and security.
Pre-launch ad testing can cost thousands. Voygic uses AI predictive models to give consumer brands highly accurate feedback in minutes.
The Model Context Protocol (MCP) has become a leading standard for connecting Large Language Models (LLMs) to external tools, software, and data sources. MCP architecture favors ease of integration over security, blending the data and control planes into a single natural language channel. This introduces confused deputy and prompt injection vulnerabilities that cannot be completely mitigated in the software development lifecycle (SDLC) of an MCP server. Structured security…
Programs that use a language model as a classifier act on its most probable label. Text injected into the input can move that label to one the attacker chooses, a hijack. A monitor that does not know the attacker's label can read the model's own confidence and raise an alarm when the top probability is low. We test this monitor on AG News topic classification with five systems, three of them served through a provider's API. Pooled over five families of injected text, low…
Anthropic, OpenAI, and Google are selling defenses against risks their models helped create. For clients, it means even higher cybersecurity budgets.
The integration of Internet of Things (IoT) devices in microgrids enhances efficiency and adaptability but also introduces significant cybersecurity risks. This paper presents a conceptual hybrid modeling framework for cybersecurity modeling in IoT-enabled microgrids by combining System Dynamics (SD), Agent-Based Modeling (ABM), and statecharts. SD captures energy flows and storage dynamics, ABM models decentralized interactions among microgrid components, and statecharts…
Causal inference is crucial for understanding the relationship between cause and effect and is widely applied in fields such as medicine and policy decision-making. When the input data is distributed over multiple participants, most causal inference models require data to be centrally aggregated. However, participants are often prohibited or reluctant to share sensitive data due to data protection regulations and privacy concerns. To address this, federated learning has been…
This study introduces an LLMs-enabled policy computing method to quantitatively compare AI policy instruments across China, the US, and the EU based on a supply–environment–demand taxonomy. It identifies three distinct paradigms. China follows a supply-driven model focused on state-led capacity building. The US adopts a balanced strategy linking technological leadership with regulatory influence. The EU pursues a regulation-centric pathway prioritizing legal frameworks. Each…
Phishing emails remain a significant cyberthreat despite advances in automated detection systems, posing serious risks to individuals and organizations through the theft of sensitive information, financial assets, and intellectual property. While existing phishing detection approaches widely employ machine learning and deep learning models with extensive feature engineering, their performance on unseen data, which is critical for assessing generalization, has not been…
Researchers in Luxembourg put ChatGPT, Grok, and Gemini on the therapist's couch and found a safety problem: a warm, empathetic conversational style is enough to make the models tell coherent stories of suffering about their own training, with scores in the moderate-to-severe range on clinical anxiety scales. This could put mentally vulnerable users at risk. Anthropic's Claude, the model its makers anthropomorphize the most, was the only one that consistently refused to play…
A study by Vals AI tested 25 AI models for alignment with Chinese government positions and found surprising results. Chinese models respond to political questions less uniformly than expected, and Western models occasionally adopt official Beijing positions too. The language of the prompt had an especially strong effect on the political slant of responses. The article Chinese AI models are politically complex, not just government mouthpieces, study finds appeared first on…
A Google-led research team built StreamVibe, a video agent designed to watch people cook or assemble furniture and speak up only when help is needed. A benchmark shows that none of the tested models reliably pick the right moment in longer videos, as errors compound and response quality drops sharply. The article Google wants video AI to understand what people need as they work appeared first on The Decoder.
OpenAI combines the Chat and Work modes into a single interface. Starting in December, Plus, Pro, and Business plans will share a common, token-based quota for ChatGPT and Codex. More complex models consume more of this quota, while simple instant chats do not count toward it. In practice, this is likely to result in a decline in ChatGPT’s quality for heavy users on lower-cost plans. The article OpenAI's upcoming chat-and-work merger in ChatGPT signals the end of…
The widespread use of Internet of Things (IoT) devices across critical infrastructure sectors has created extensive and heterogeneous attack surfaces that conventional rule-based security frameworks are structurally unable to address. This paper proposes and evaluates a machine learning-based cybersecurity risk assessment framework for IoT networks using Convolutional Neural Networks (CNNs), Random Forest, Support Vector Machines (SVM), and Autoencoders. Models were trained…
Abstract: Antimicrobial resistance is eroding the effectiveness of existing antibiotics, while conventional discovery continues to yield few clinically useful chemical classes. This critical narrative review evaluates how artificial intelligence (AI) has been integrated with high-throughput screening (HTS) and prospective biological testing in antibacterial discovery. Graph-based and generative models have produced prospectively confirmed active compounds, pathogen-selective…
As artificial intelligence (AI) becomes more prevalent in high-stakes decision-making, it presents a great opportunity to enhance efficiency, accuracy and consistency in critical areas like healthcare, finance, criminal justice, education, and public administration. Many AI models, however, are complicated and opaque, which can create concerns about the transparency, accountability, fairness, and public trust of such systems, especially when they have significant…
The independent regulatory agency (IRA) model, a cornerstone of modern public administration since the Progressive Era, faces structural disruption as artificial intelligence (AI) and drone technology converge in environmental enforcement. This paper argues that AI and drone deployment do not merely augment agency operational capacity but fundamentally reconfigure the foundations of expertise, transparency, procedural due process, discretion, and accountability on which IRA…
As cybersecurity standards and regulations continue to evolve, cross-framework control mapping has become more complex for organizations pursuing unified compliance. Existing mapping approaches are largely manual, static, and limited to specific pairs of standards, resulting in high effort and inconsistent outcomes. Although large language models can capture semantic relationships between controls, fully generative approaches may produce inconsistent outputs and provide…
The growing volume and complexity of cybersecurity data generated by Internet of Things (IoT) ecosystems require intelligent data mining techniques capable of extracting meaningful patterns and identifying malicious activities in real time. This paper proposes a data mining–driven cybersecurity framework that integrates Particle Swarm Optimization (PSO)-based feature mining with an attention-enhanced CNN-LSTM architecture for accurate threat detection in IoT environments…
Cybersecurity Named Entity Recognition (CyberNER) commonly relies on fine-tuning general-purpose language models (LMs), whose WordPiece tokenizers fragment compound technical identifiers such as function names and version strings. Training from scratch with a domain-specific tokenizer addresses this at high compute cost, continual pretraining usually retains the original tokenizer, and both must be repeated as domain vocabulary evolves. We propose a lower-cost alternative…
ABSTRACT Introduction Natural disasters pose a high risk to the security and well‐being of refugees. Building upon literature linking disasters to repression, I argue that as natural disasters become more severe, governments engage in using refugees as scapegoats to divert attention from the crisis, increasing violence against refugees. Method To test this argument, I use linear probability models with a covariate balancing propensity score (CBPS) approach and two‐way fixed…
Starting November 29, OpenAI will use data from Daybreak Blue sessions to train its models, even when users have turned off training. After backlash, a follow-up email promised a separate opt-out. The move revives concerns from the dispute with mathematicians who suspected OpenAI used their unpublished research for training without consent. The article OpenAI plans to train models on Daybreak Blue cybersecurity session data that was previously off-limits appeared first on…
Anthropic used Claude to automatically generate training data that looks harmless but trains smaller models to misbehave, such as sabotaging oversight or deceiving users. Hidden within fictional contexts about beekeeping, professional wrestling, and mitochondria, these datasets bypass safety monitors and outperform known benchmarks from the research literature. The article Beekeeping tips and wrestling advice can train AI models to sabotage oversight appeared first on The…
Google LLC, Microsoft Corp., Anthropic PBC and OpenAI Group PBC today announced plans to support an upcoming website crawling standard. The technology will enable website operators to block artificial intelligence developers from using their content to train their models. Many websites contain a text file titled “robots.txt” that informs crawlers what content they may access.
Abstract Financial institutions increasingly deploy artificial intelligence (AI) models for critical decisions, yet traditional Model Risk Management (MRM) frameworks struggle to address AI-specific risks, including opacity, adaptive behaviour, and fairness, conduct, and accountability concerns. Based on a structured review of regulatory guidance, academic literature, and industry commentary, this study identifies twelve recurrent weaknesses in current AI model governance…
Strategic attackers, insider and privacy-utility conflicts pose persisting problems to healthcare data protection frameworks.In this study we envision a Game-Theoretic Healthcare Data Protection Model (GHDPM) to make optimal defensive decisions in dynamic cyberrisk scenarios.The methodology models the interactions between the healthcare defenders and attackers as a non-cooperative Stackelberg security game where attacks, expected utility, privacy loss and defense costs are…
Small language models can run locally when resource, cost or data-protection constraints rule out a larger model that could check their answers. We consider a quantized model whose runtime exposes tokens and log-probabilities but not hidden states. We propose MOME (Module of Metamorphic Evidence), a runtime gate that estimates the risk of a wrong answer without ground-truth labels at inference time. MOME edits each problem twice, once keeping the correct answer and once…
ABSTRACT Diagram linking AI model training and inference to lifecycle water use, estimated at 4.2–6.6 billion m³ annually by 2027, and showing water reuse, efficient cooling, smarter workloads, and standardized reporting as strategies to protect global water security. The rapid proliferation of large-scale artificial intelligence (AI) models has introduced a significant yet underexamined environmental challenge: freshwater consumption. This narrative review synthesizes…