News about foundation models, LLMs, multimodal models, benchmarks, and model releases from OpenAI, Anthropic, Google, Meta, Mistral, and others.
973 articles
According to the Financial Times, hackers are increasingly hijacking corporate AI model and cloud accounts, then reselling access on dark web marketplaces. The trend, dubbed 'LLM jacking,' highlights growing security risks tied to enterprise AI adoption. Source: biz.chosun.com
Discussions about AI security tend to focus on the model itself — whether it's aligned, jailbreak-resistant, or prone to hallucination. But a growing risk lies in the disconnect between what AI agents can read and what systems they're allowed to actually modify. Source: venturebeat.com
KT built an LLM router that analyzes a user's query and automatically selects the most suitable AI model to answer it. The system, called Router Arena, benchmarks and compares different LLMs to route requests efficiently. Source: mk.co.kr
KT's in-house tool AutoModelRouter, which automatically selects and routes queries among various AI models, took second place in a public benchmark of LLM routers. The result highlights growing competition around efficient model-selection technology. Source: biz.chosun.com
TypeSafe AI's Jev was tested across 3,080 classification tasks to compare its accuracy, latency, calibration, and confidence against standard LLMs. The goal was to see whether a decision-focused AI system can outperform generative models on structured tasks. Source: towardsdatascience.com
Boost Mobile data scientist Priyank Jain says the telecom industry underestimates unglamorous data issues that quietly determine whether AI models keep working correctly. He argues model skew deserves far more attention than it currently gets. Source: fiercewireless.com
Researchers are increasingly testing large language models for tasks like interpreting scientific literature, extracting data, and supporting statistical analysis in evidence synthesis. The study examines how reliably LLMs can handle automated meta-analysis generation. Source: nature.com
Government agencies could save money and gain better oversight by choosing model size based on the actual task rather than automatically picking the largest available model. Smaller, purpose-fit models can offer comparable results with lower cost and risk. Source: fedtechmagazine.com
IBM describes large language models as advanced AI systems trained on massive datasets to understand and generate human-like text. The piece looks at how their widespread adoption is reshaping communication, work, and society more broadly. Source: daily-sun.com
Researchers introduced TimeBraid, a family of models that align pretrained language models with pretrained time-series foundation models. The goal is to combine strengths of both approaches for better understanding and forecasting of temporal data. Source: alphaxiv.org
Anthropic announced that its Claude model helped identify a gene-editing system resembling CRISPR. Commentators caution this is unlikely to be a revolutionary scientific leap, more likely just another useful addition to the gene-editing toolbox. Source: newscientist.com
LLMs are increasingly used to offer text-based recommendations in fields like urban planning, design and public health. The growing role raises questions about whether these models can be trusted to provide sound ethical guidance in areas affecting people's wellbeing. Source: techxplore.com
Federal agencies don't always need the largest available AI model — picking a size that matches the actual task can lower costs and improve oversight. Experts argue this approach offers a more practical path to responsible AI adoption in government. Source: fedtechmagazine.com
Traditional herbal medicine has built up a vast but scattered body of knowledge about which plants may treat which illnesses. Language models are now being applied to sift through this sparse, fragmented historical data and reveal previously hidden herb-disease connections. Source: bioengineer.org
Taiwan's Ministry of Digital Affairs released evaluation results assessing how well AI models genuinely understand Taiwan, not just its language. The initiative aims to bring government and industry together to build a trustworthy, locally grounded AI ecosystem. Source: moda.gov.tw
Liner has launched a new API that automatically directs each request to the cheapest model capable of handling it, rather than always using expensive frontier models. The company says this approach can cut enterprise LLM costs by more than half. Source: manilatimes.net
Researchers found that a large language model can detect cyberattacks on smart power grids more effectively than conventional security defenses. The finding matters as energy infrastructure remains one of the most frequently targeted digital systems worldwide. Source: bioengineer.org
A former DeepMind researcher is reportedly raising tens of millions of dollars for a new venture called Metis Reasoning. The move signals growing investor interest in AI approaches that go beyond traditional large language models. Source: pymnts.com
A Google engineer named Patrick suggests that Anthropic's Opus 5.5 may have been created through recursive self-improvement, possibly distilled from a much more powerful internal model referred to as Model 2. The claim remains an unconfirmed hypothesis rather than an official statement from Anthropic. Source: eu.36kr.com
NVIDIA is pushing confidential computing as a way to protect sensitive data and proprietary model context during large language model inference. The technology targets enterprise and personal use cases where privacy and security are critical as LLM deployment scales up. Source: developer.nvidia.com