Claude Mythos Preview's Security Bombshell, AlphaGenome Decodes Dark DNA, AI as a Mirror for the Blind, and Walrus Simulates Fluids

Published on 10.04.2026

AI & AGENTS

Claude Mythos Preview: Anthropic's Security Warning Without a Product Launch

TLDR: Anthropic published a 244-page model card for Claude Mythos Preview, a model so capable at finding real-world vulnerabilities that they assembled a 40+ organization security consortium before disclosing it. The model is not commercially available.

The Batch — deeplearning.ai


AlphaGenome Makes the "Junk DNA" Problem Tractable

TLDR: Google DeepMind released AlphaGenome, an open-weights model that interprets the 98% of the human and mouse genomes that don't code for proteins, predicting gene expression properties and mutation effects with state-of-the-art results across 50 evaluations.

The Batch — deeplearning.ai


The AI Mirror Problem: When Vision Models Judge Blind Users

TLDR: A blind journalist's account of using GPT-4 Vision as a mirror raises substantive questions about what AI vision apps designed for blind users are actually optimizing for — and whose aesthetic standards they encode.

The Batch — deeplearning.ai


Walrus: A General Transformer for Fluid Dynamics Across 19 Physical Domains

TLDR: Polymathic AI released Walrus, a 1.3B-parameter transformer that simulates liquids, gases, and plasmas across 19 physical domains including acoustics, astrophysics, and non-Newtonian fluids. It matches or exceeds specialized physics models and is MIT licensed.

The Batch — deeplearning.ai