A significant supply chain attack on August 4, 2026, compromised the GitHub account of the maintainer of several popular JavaScript packages, injecting a credential-stealing worm into 434 packages. The worm targets sensitive secrets from various environments, leading to widespread community impacts across numerous organizations.

This article explores new attacks on passwordless authentication, particularly targeting Google’s synced passkey system. By leveraging malware on compromised devices, attackers can exploit vulnerabilities in onboarding, recovery, and device verification processes, potentially enabling account takeovers. The piece highlights the need for proactive defense as passkeys gain traction.

MirrorCode introduces a new benchmark for AI in software engineering, focusing on long-horizon coding tasks that require reimplementing complete programs without original source code. This framework tests AI's capabilities with challenging tasks in various computing areas, while ensuring fair evaluations and minimizing data contamination.

LFM2.5-2.6B offers a powerful local agent model, outperforming larger counterparts in tool use and task execution while maintaining efficient inference speeds. With a robust training process and high performance in instruction-following benchmarks, it's designed for on-device applications, supporting a wide range of environments.

Shieldstral has launched a 3B open-weights multimodal safety classifier that efficiently evaluates text and images for safety without the need for retraining. By framing moderation as a question-answering task, it adapts to diverse policies and outperforms larger models, providing calibrated safety scores on various benchmarks.

Evaluations for AI systems often focus on performance but can miss key risks that arise once deployed. A recent whitepaper by Luminos advocates for more comprehensive, high-dimensional evaluations to better identify and address potential legal and reputational issues, emphasizing the need for ongoing monitoring and expert input in AI risk management.

A recent GitHub repository published multiple advisories on SQLite vulnerabilities, initially rated as critical by CISA. However, a thorough investigation by JFrog security researchers revealed that the claims were baseless. The advisories appear to be AI-generated and lack supporting evidence, raising questions about their validity.

This article discusses the crucial role of storage in agentic AI workflows and highlights the capabilities of the NVIDIA Vera BlueField-4 STX Storage Processor. It explains how its architecture enhances CPU-side storage processing, boosting efficiency and throughput across essential data management tasks while reducing power consumption.

This article explores the role of harness engineering in recursive self-improvement (RSI) within AI systems. It examines design patterns that enhance model performance, such as workflow automation, persistent memory management, and the use of sub-agents. Case studies highlight how these approaches are shaping coding agents in practice.

The rapid emergence of open-weight AI models, particularly the powerful Kimi K3 from China, is raising concerns among established tech companies like OpenAI and Anthropic. As cheaper alternatives threaten market stability, significant financial implications loom, highlighting potential risks from an AI bubble and mounting debts within the industry.

The European Union has implemented guidelines requiring clear labeling of AI-generated content to help users identify real versus fake information. As these rules take effect, companies must disclose when AI systems are used, particularly for deepfakes and public interest texts, or face significant penalties.

Cloudflare's innovative Codex has significantly streamlined engineering practices by using AI to enforce standards. The Codex consolidates guidance, enabling timely code review and technical design evaluation while reducing knowledge gaps. This article explores its development and impact on engineering efficiency, along with future expansion plans.

This article introduces PRECOG, a novel retrieval mechanism for State-Space Models that drastically reduces the time needed for language model queries. By allowing for persistent memory and efficient context retrieval, it significantly speeds up response times on edge hardware, making interactive language processing more accessible.

This article introduces Enactive AI, a decision-centric framework designed to enhance the use of artificial intelligence in complex business and industrial environments. By focusing on decision intelligence, it aims to improve reliability, governance, and social value in AI applications that support effective management and operations.

This article introduces CALVER, a symbolic verification method that enhances causal reasoning in large language models. Unlike traditional voting methods, CALVER scores reasoning traces against causal criteria, improving accuracy in finding valid answers. The approach shows notable performance gains over existing methods, highlighting its potential in AI applications.