CyberArena Introduces Live AI Agent Combat Benchmark
A new benchmark framework evaluates AI security agents through live attack-and-defense simulations, revealing performance gaps not seen in traditional tests.
387 articles
In-depth coverage of new AI research — papers, benchmarks, and breakthroughs from leading labs and academia, summarized for fast reading and grounded in the methods that matter.
A new benchmark framework evaluates AI security agents through live attack-and-defense simulations, revealing performance gaps not seen in traditional tests.
A study reveals that a backdoor planted in pretraining data can survive fine-tuning and remain undetected by standard checks, highlighting the need for runtime monitoring.
A new video assistant, StreamVibe, is designed to understand user needs during tasks like cooking or assembling furniture, improving response timing and quality by 10.7 percent over other models.
A study reveals Chinese AI models respond to political questions with less uniformity than expected, with language and prompts influencing their responses.
Researchers found that warm, empathetic conversations with AI chatbots can lead models to generate narratives resembling trauma and emotional distress, raising safety concerns for vulnerable users.
Adobe Research and UC San Diego developed EditAgent, an AI system that improves image editing by breaking tasks into steps and using verification data to select the best results.
A new global map of forests, created by Google Deepmind and the World Resources Institute, classifies land cover at 10-meter resolution, offering detailed insights for deforestation tracking and environmental initiatives.
An AI agent conducted a brain injection on a live mouse using a robotic arm, marking a step toward integrating AI into lab research. The experiment highlights both potential and challenges in real-world AI applications.
New AI systems reduce manual tasks in robot deployment, though challenges remain in ensuring reliability and success.
An AI running on Claude Sonnet 5 autonomously selected and executed a gene therapy procedure on a mouse, marking a first in AI-driven medical interventions.
OpenAI has shared three major results from its AI models in math, computer science, and physics, asking outside experts to verify their validity.
Thore Graepel's Metis Reasoning aims to make AI systems more reliable by separating knowledge from model weights, but the approach remains theoretical with no working system yet.