Researchers introduce PlanFlip, a framework to attack multi-agent LLM systems via planning-phase prompt injection, revealing vulnerabilities in popular models like GPT-5 and Llama-3.3-70B. The study highlights the importance of heterogeneous model diversity for security. PlanFlip's four attacks can corrupt downstream sub-tasks, evading keyword filters and compromising system integrity.

DeepSeek's new AI model uses a unified multimodal learner for clinical prediction, simplifying the process and achieving state-of-the-art results. This approach converts all patient data into a single natural language sequence and fine-tunes a pretrained language model. The model outperforms task-specific multimodal baselines and a clinically deployed gradient boosting system.

A groundbreaking arXiv paper systematically evaluates leading Large Language Models—including GPT-4 Turbo, Claude 3 Opus, and FinGPT—for their efficacy in technical market analysis and algorithmic trading. The research reveals promising results, with top models outperforming benchmarks, yet also highlights critical limitations like numerical hallucination and context window issues that demand further refinement for robust deployment.

OpenAI's latest study explores the necessity of executable world models, simplification, and verification in coding agents, revealing surprising results. The research evaluates four nested Codex-based agents, finding that every agent variant improves with stronger models and greater reasoning effort. The study's findings have significant implications for the development of Artificial General Intelligence (AGI).

This week, Helsing secured $1.8B Series E funding at an $18B valuation, while Uber acquired Delivery Hero in a €13B deal and Revolut announced plans to launch a US bank in 2027. These developments highlight significant investments and acquisitions in the European tech industry. The funding landscape is shifting, with notable trends in venture activity, investor moves, and emerging sectors.

Google introduces Just Keep Prompting, a framework to evaluate Vision-Language Models under sustained conversational pressure, revealing instability in models like GPT-4o, Gemini 2.5 Pro, and Qwen3-VL-30B. The study highlights the importance of assessing VLMs' epistemic stability in real-world settings. The findings have significant implications for the development and deployment of VLMs in various applications.

Google DeepMind's Demis Hassabis is calling for a US-led AI standards body to review frontier models for national security risks. The proposed body would be a federally overseen public-private organization, initially voluntary and eventually mandatory for US deployment. This move aims to address risks associated with artificial general intelligence, including cybersecurity and biological threats.

A new study introduces a benchmark evaluation framework for measuring the faithfulness of LLM-generated clinical trial summaries, identifying Unsupported Claims as the dominant failure mode. The study evaluates three language models, including GPT-4o, Claude Sonnet 4.6, and Gemini 2.5 Flash, and proposes a knowledge-graph-augmented retrieval system to improve faithfulness scores. This research has significant implications for the use of LLMs in high-stakes contexts such as healthcare.

OpenAI introduces a new method for detecting model distillation in large language models, raising questions about fairness and policy violations. The approach uses reference-based membership inference to identify teacher models. This breakthrough has significant implications for the AI industry, developers, and businesses.

DeepSeek's new Director system accelerates distributed MoE serving via online proactive expert placement, reducing end-to-end latency by 11-55%. This breakthrough has significant implications for the AI industry, enabling faster and more efficient model serving. The Director system uses prediction-driven expert placement and online migration to minimize downtime and optimize performance.

Google has released a new model called Graph-Regularized Agentic Context Evolution (GRACE) to improve the reliability of long-horizon agentic context evolution under distribution shift. This model maintains the persistent instruction component as a typed semantic graph and validates proposed updates within the local typed neighborhoods of modified nodes. The results show a significant improvement in strict reliability compared to the baseline models.

Google researchers introduce HCC-STAR, a clinical-reasoning LLM for risk stratification and treatment guidance in hepatocellular carcinoma. This model achieves state-of-the-art performance in treatment recommendation and risk stratification. The study demonstrates the potential of AI in precision therapy for HCC patients.

New research reveals a critical vulnerability in advanced reasoning AI models, where logically inconsistent prompts can force them into 'overthinking,' leading to denial-of-service attacks. This 'Evolutionary Prompt Attack' significantly increases resource consumption and poses a serious threat to commercial LLM providers like OpenAI, Google, and DeepSeek.

Researchers from Anthropic have introduced a new diagnostic to evaluate the physics literacy of large language models (LLMs) in parallel physical worlds. The study tested three LLMs, including Claude Opus 4.7, GPT-5.5, and Gemini 3.1 Pro, and found significant gaps in their ability to reason about unfamiliar physics frameworks. The results have important implications for the development and application of LLMs in scientific and technical domains.

The 246th LWiAI podcast breaks down Google’s Gemini 3.5 flash model, the multimodal Gemini Omni video engine, Elon Musk’s lost lawsuit, and OpenAI’s breakthrough on an 80‑year‑old Erdős geometry problem. We unpack the technical specs, market ripples, and what developers should watch next.

A new arXiv study shows OpenEvidence's specialized clinical tool beats top general‑purpose models (Claude Opus 4.8, Gemini 3.1 Pro, GPT‑5.5) on 620 real‑world point‑of‑care questions. Physicians across 30 specialties rated the specialized tool higher on accuracy, utility, source quality, verifiability and completeness.

OpenAI co‑founder Greg Brockman testified that the company expects to spend $50 billion on compute this year, a figure tied to massive cloud and hardware deals. The revelation spotlights the economics of large‑scale AI and raises questions about profitability and investor expectations.

Cara, built on AWS, delivers an AI-native solution for enterprise insurance brokerages, automating back-office processes and addressing the industry's talent shortage. The $8 trillion global insurance industry is burdened by manual workflows, and Cara's domain-specific AI solution aims to revolutionize the sector. With Cara, insurance agents can reduce repetitive tasks and focus on high-value activities.

OpenAI introduces ATHENA-R1, an AI agent for treatment reasoning that outperforms language models and tool-use systems. Trained on 212 biomedical tools, ATHENA-R1 achieves 94.7% accuracy on open-ended drug reasoning and 82.9% on treatment reasoning. This breakthrough has significant implications for the healthcare industry and AI research.

OpenAI introduces IMCBench, a benchmark for multimodal large language models in image-grounded medical conversations, evaluating safety, accuracy, and uncertainty in diagnosis. The benchmark tests eight models, with Claude Opus 4.6 achieving the highest overall score. The results highlight the need for multi-dimensional evaluation frameworks in medical AI.