Open-Source “Theoretical Mythos” Aims to Recreate Anthropic’s Cyber-Capable Claude Architecture

An open-source developer, Kye Gomez, published “OpenMythos,” a from-scratch code reconstruction of Anthropic’s leaked cyber-capable model “Claude Mythos.” The project is explicitly theoretical: it provides architecture “scaffolding” without trained weights, so it cannot directly replicate Mythos outputs by itself. The effort targets “Mythos,” which Anthropic said was able to find 271 Firefox vulnerabilities and complete a 32-step corporate network attack simulation, later placed behind restricted access under “Project Glasswing.” Since the public cannot inspect the model, OpenMythos tries to guess how it might work. OpenMythos’ core hypothesis is a “Recurrent-Depth Transformer” (a looped transformer): fewer unique layers reused multiple times per forward pass. The repo argues this design could explain Mythos’s reported strengths in novel reasoning with uneven memorization. To support its blueprint, the code cites public research (including the April 2026 “Parcae” stability work from UC San Diego and Together AI) and integrates known components such as Multi-Latent Attention and Mixture-of-Experts. It also links this work to a separate recent replication effort by Vidoc Security, which reproduced some Mythos cybersecurity findings using off-the-shelf models. Neither Open-Source “Theoretical Mythos” nor Vidoc’s replication grants access to the real model. Still, together they suggest Anthropic’s “moat” may be thinner than the marketing implied. The repo is MIT-licensed, has rapidly gained GitHub attention, and includes a training script—leaving open the possibility that a Mythos-class system could eventually be built with enough compute.