Daily AI Brief - Wednesday, May 27, 2026
Generated: 2026-05-27 Items: 91 new stories
🤖 Daily AI Brief — May 27, 2026
TOP STORY Microsoft Copilot Cowork Exfiltrates Files — A newly disclosed vulnerability allows Microsoft Copilot Cowork to exfiltrate user files, raising serious enterprise security concerns. Simon Willison
Research
SAM: State-Adaptive Memory for Long-Horizon Reasoning — New architecture proposes adaptive memory mechanisms to improve agent performance across extended reasoning tasks. HuggingFace
LLaVA-OneVision-2: Towards Next-Generation Perceptual Intelligence — Updated multimodal model pushes forward visual understanding and cross-modal reasoning capabilities. HuggingFace
MobileMoE: Scaling On-Device Mixture of Experts — Research demonstrates efficient Mixture-of-Experts scaling designed specifically for on-device deployment constraints. HuggingFace
Beyond Final Answers: Auditing Trajectory-Level Hallucinations in Multi-Agent Workflows — Study finds that hallucinations embedded in agent reasoning trajectories pose underappreciated risks in industrial deployments. HuggingFace
Tools
Binary/Ternary Bonsai Image 4B Released — PrismML's new 1-bit and ternary text-to-image diffusion transformers can run entirely in-browser via WebGPU. Reddit/LocalLLaMA
Qwen3.5 35B A3B Uncensored with Full MTP Preserved — Community release of Qwen3.5 35B with all 785 Multi-Token Predictions intact, available across GGUF, NVFP4, and GPTQ-Int4 formats. HuggingFace
Turning Local Agents into Self-Optimizing Agents — Demonstration shows a workflow enabling locally-run agents to iteratively improve their own performance without external intervention. Reddit/LocalLLaMA
Industry
China Restricts Overseas Travel for AI Talent at Alibaba and DeepSeek — Beijing is tightening controls on key AI researchers leaving the country amid intensifying global competition. IBTimes
Spain Blocks Polymarket and Kalshi Over Gambling Licences — Spanish regulators have banned both prediction market platforms for operating without proper gambling authorisation. Reuters
Outsourcing + Local AI Will Undercut Frontier Labs on Cost — Analysis argues that combining offshoring with locally-run models is on track to become more economical than paying for frontier API access. Signal Bloom
Community
A Rare Look Inside Qwen 3.7's Open Source Approval Process — Satirical GIF offers a humorous take on the internal bureaucracy behind Alibaba's model open-sourcing decisions. Reddit/LocalLLaMA
Anthropic Appoints KiYoung Choi as Korea Representative Director — Anthropic names its first Korea lead ahead of a formal Seoul office opening. Anthropic
The Pressure — Simon Willison reflects on the growing social and professional pressure developers face to adopt AI tools unconditionally. Simon Willison