May 22, 2026

• Antigravity 2.0 Tops the OpenSCAD Architectural 3D LLM Benchmark
  The OpenSCAD LLM benchmark evaluated multiple AI coding tools on their ability to generate a detailed Pantheon model, revealing significant differences in output quality and speed among Codex 5.5 High, Claude Sonnet, and Google Antigravity 2.0, with Antigravity achieving the best autonomous result.

• Project Glasswing: An Initial Update
  Project Glasswing has identified over 10,000 critical vulnerabilities in essential software using Claude Mythos Preview, significantly enhancing the speed of vulnerability detection compared to traditional methods.

• The current AI pricing was always going to go away
  AI pricing is shifting dramatically as companies like Microsoft and GitHub abandon flat-rate models due to rising costs in memory and GPU resources, which have surged by 4x and 95%, respectively, in recent months.

• ACC: Compiling Agent Trajectories for Long-Context Training
  Agent Context Compilation (ACC) transforms agent-generated trajectories into long-context QA pairs, enabling LLMs to integrate scattered evidence from multiple turns without additional annotation, thus enhancing reasoning capabilities.

• NuExtract3 released: open-weight 4B VLM for Markdown, OCR and structured extraction (self-hostable)
  NuExtract3 is a 4B model designed for information extraction from complex documents, including PDFs and forms, and is available under the Apache-2.0 license.

• Domain-Camouflaged Injection Attacks Evade Detection in Multi-Agent LLM Systems
  Domain camouflaged injection attacks exploit the vocabulary and authority structures of target documents, leading to a dramatic drop in detection rates from 93.8% to 9.7% on Llama 3.1 8B and from 100% to 55.6% on Gemini 2.0 Flash, revealing a critical vulnerability in current detection systems.

• Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention
  Gated DeltaNet-2 innovatively separates the roles of erasing and writing in linear attention, utilizing channel-wise gates to enhance memory editing without compromising existing associations.

• Can liveness detection models generalise to synthetic media generation techniques they were never trained on?
  Liveness detection models may struggle to generalize to new synthetic media generation techniques, as they were primarily trained on outdated datasets that do not reflect current advancements in deepfake technology.

• Live Human Detector on Outbound Phone Calls
  The Live Human Detector aims to enhance call center efficiency by accurately identifying when a call transitions from an automated system to a live agent within a 1-2 second window, utilizing advanced audio classification techniques.

• Vega: Zero-knowledge proofs for digital identity in the age of AI
  Vega enables users to prove facts from government-issued credentials, such as age or professional status, without revealing the credential itself, ensuring privacy and security. This is achieved through zero-knowledge proofs generated in under 100 ms on standard devices, making it scalable for real-world applications like mobile driver’s licenses and the EU Digital Identity Wallet.

• Microsoft Drops Claude Code After Budget Overrun
  Microsoft's Claude Code pilot consumed its entire annual AI budget in mere months due to token-based billing, leading to a cancellation effective June 30, 2026, and redirecting developers to GitHub Copilot.

• NVIDIA GTC Taipei at COMPUTEX: Live Updates on What’s Next in AI
  NVIDIA GTC Taipei at COMPUTEX showcases cutting-edge advancements in AI, featuring live demos and a keynote by CEO Jensen Huang on June 1, 2026, emphasizing the convergence of developers and industry leaders to explore innovations in AI factories and autonomous systems.

• MagenticLite, MagenticBrain, Fara1.5: An agentic experience optimized for small models
  MagenticLite integrates MagenticBrain and Fara1.5 to create a seamless agentic experience optimized for small models, enabling efficient task execution across browsers and local systems.

• Masked Diffusion Language Models are Strong and Steerable Text-Based World Models for Agentic RL

• DeferMem: Query-Time Evidence Distillation via Reinforcement Learning for Long-Term Memory QA
  DeferMem introduces a novel long-term memory framework that enhances query-time evidence distillation through a reinforcement learning approach, effectively organizing and retrieving relevant information from extensive conversational histories.