AI News — Wednesday, August 19, 2026

9
StateM: Achieving 95.3% Raw Accuracy on Terminal-Bench 2.1 with Harness Scaling

A new paper introduces StateM, a method that achieves a high raw accuracy of 95.3% on the Terminal-Bench 2.1 benchmark, demonstrating significant progress in AI efficiency assessment.

Hugging Faceresearch
9
OpenAI Institutes New Safeguards Following Hugging Face Breach

OpenAI has implemented enhanced security measures after a recent breach at Hugging Face, highlighting ongoing concerns about AI platform security and data integrity.

TechCrunchindustry
8
Anthropic Launches Cowork: A Claude Desktop Agent for File Interaction Without Coding

Anthropic has released Cowork, a new Claude desktop agent designed to interact directly with user files and perform tasks without requiring any coding, aiming to make AI more accessible for everyday productivity.

VentureBeatproduct
8
Etched's Valuation Soars to $21 Billion in One Month

AI hardware startup Etched has seen its valuation double to an astonishing $21 billion within a single month, signaling intense investor confidence and rapid growth in the specialized AI chip market.

TechCrunchindustry
7
Cursor Launches Rival Hosting Platform, Capitalizing on GitHub Frustration

Cursor has introduced a new code hosting platform, positioning itself as a direct competitor to GitHub by addressing developer frustrations and aiming to attract users with AI-enhanced features.

TechCrunchproduct
7
HarnessEval-W: A New Framework for Agentifying the Evaluation of Visual Worlds

Researchers introduce HarnessEval-W, a novel evaluation framework designed to 'agentify' the assessment of visual worlds, providing a more dynamic and interactive approach to benchmarking AI agents.

Hugging Faceresearch
7
Amazon AI Training Facility Tracked Receiving Shipment of Rare Books

An investigation revealed that a shipment of rare books was tracked to an Amazon AI training facility, raising questions about data acquisition practices for large language models.

Lobste.rsindustry
7
OpenAI Blog Post: Pacing Model Development in an Era of Cyber-Critical Capabilities

OpenAI discusses the critical need for careful pacing in AI model development, especially as capabilities become increasingly relevant to cybersecurity and national security.

OpenAI Blogindustry
6
VibeWorlding: Multimodal Agents Construct 3D Open Worlds End-to-End

A new research paper explores the capability of multimodal AI agents to construct complex 3D open worlds from end-to-end, pushing the boundaries of generative AI in virtual environments.

Hugging Faceresearch
6
COSP: A Prompting Trick Enabling LLMs to Grade Their Own Homework

COSP introduces a novel prompting technique that allows Large Language Models to self-evaluate and 'grade their own homework,' potentially improving their learning and refinement processes.

Dev.toresearch
6
Large Discovery Models: Empirically-grounded Model-Based Open-Ended Search

This paper introduces Large Discovery Models (LDMs) that leverage empirically-grounded, model-based open-ended search to explore and generate novel solutions or data.

Hugging Faceresearch
5
Learn What's Left, Not What's Mastered: Saturation Aware Advantage Reweighting for Multi-Reward Policy Optimization

A new method called Saturation Aware Advantage Reweighting is proposed to optimize multi-reward policies by focusing learning on less-mastered aspects, improving efficiency in complex reinforcement learning tasks.

Hugging Faceresearch
5
UI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations

UI-Mate presents a significant step forward for open-weight foundation GUI agents by enhancing their capabilities through in-context demonstrations, making them more adaptable to various user interfaces.

Hugging Faceopen-source
5
MOSS-VL Technical Report: A New Multimodal Open-Source Vision-Language Model

The MOSS-VL Technical Report details a new multimodal open-source vision-language model, contributing to the growing ecosystem of publicly available advanced AI models.

Hugging Faceopen-source
5
OpenAI Partners with CodeAI to Prepare the First AI Generation

OpenAI has announced a partnership with CodeAI, focusing on educational initiatives to prepare the next generation for an AI-driven future, emphasizing skills and understanding in artificial intelligence.

OpenAI Blogindustry