Microsoft: AI Costs Higher Than Human EmployeesLatent Space Attacks Bypass AI Safety ControlsIs AI Profitable Yet? Industry AnalysisMindLoom: New Method for High-Quality Reasoning DataNew Method Tackles LLM Benchmark Data ContaminationTrace2Skill: Boosts Long-Context EDA Agent PerformanceSMDD-Bench: Evaluates LLM Drug Design CapabilitiesCenterLoss Hurts OOD Detection, Multi-Scale Mahalanobis WinsAll Model Labs Renamed to Agent LabsClaude Code v2.1.150 ReleasedOpenClaw 2026.5.22 ReleasedMicrosoft: AI Costs Higher Than Human Employees

New research shows AI safety refusals can be suppressed by manipulating internal representations. The study demonstrates latent space attacks that bypass safety controls, highlighting the need for stronger AI safety defenses.

Analysis of AI profitability reveals challenges despite potential in certain areas. High deployment costs, complex integration, and uncertain ROI remain major obstacles. Businesses need more practical AI implementation strategies to achieve commercial viability.

MindLoom introduces a novel approach for synthesizing high-quality reasoning data by composing thought modes. The system identifies structural factors affecting problem difficulty, addressing limitations of existing methods and advancing LLM complex reasoning capabilities.

Researchers develop provable joint decontamination method to address LLM benchmark data contamination. The approach accurately identifies and removes contaminated evaluation data from training sets, ensuring reliable cross-model performance comparisons.

Trace2Skill introduces verifier-guided skill evolution for complex Verilog design problems. The method enables precise localization and modification in large code repositories, significantly improving AI capabilities in hardware design.

SMDD-Bench benchmark evaluates LLM performance on real-world small molecule drug design tasks across diverse chemistries and targets. The new benchmark addresses evaluation gaps and provides critical insights for AI applications in drug discovery.

New research reveals CenterLoss harms OOD detection in ML systems. Current methods optimize features solely for classification accuracy, neglecting OOD detection capability. The arXiv paper proposes multi-scale Mahalanobis distance as a superior alternative, demonstrating better performance in identifying out-of-distribution data while maintaining classification accuracy.

All Model Labs has officially rebranded to Agent Labs, shifting its focus from single model development to comprehensive AI agent solutions. The rename aims to integrate technical resources and enhance AI agent capabilities across multiple scenarios, including task automation, cross-platform collaboration, and personalized services. The new platform supports over 50 third-party models and processes over 1 million agent requests daily, a 200% year-over-year increase. This transformation improves product flexibility and offers developers a more efficient toolchain, expected to further drive the adoption of AI agents in the enterprise market.

Claude Code releases v2.1.150 with internal infrastructure improvements, no user-facing changes. This update enhances system stability and performance, setting the foundation for future feature iterations.

OpenClaw releases 2026.5.22 with performance optimizations including gateway improvements, process-stable channel reuse, and CPU profile rotation. These updates reduce resource consumption and enhance large-scale deployment performance.