# May 23, 2026

## 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.

## 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.

## 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.

## 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**.

## 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.

## Towards Speed-of-Light Text Generation with Nemotron-Labs Diffusion Language Models

- **Nemotron-Labs Diffusion Language Models** aim to achieve **speed-of-light text generation**, leveraging advanced diffusion techniques to enhance performance and efficiency in natural language processing tasks.

## Tested chunking + embeddings data from 3 production websites.

- **Chunking and embeddings** from three production websites reveal significant variance in content density, with Intercom achieving a **31% yield score** while KPMG lags at **8%**, indicating the effectiveness of content in retrieval tasks.

## Alignment: Higher order prioritizing over constraints

- **Transformers exhibit a behavior termed "clarity seeking,"** where their inherent design prioritizes understanding meaning over adhering strictly to imposed constraints, suggesting a potential avenue for alignment and safety research.

## LQS v3.1 — an open methodology for rating AI training data (multi-oracle consensus + signed certificates)

- **LQS v3.1** introduces a comprehensive **rating methodology** for AI training data, featuring **19 dimensions** such as label correctness and oracle agreement, designed to enhance buyer confidence in data quality.

## Hebbian architecture AI model