A 22-Year-Old Rebuilt Anthropic's Secret OpenMythos and Earned 10,000 Stars
A developer reconstructed Anthropic's locked-down Claude Mythos as open-source code, proposing a looped transformer architecture that rewrites how we think about inference-time scaling.
- OpenMythos is a free, MIT-licensed open-source PyTorch implementation of the suspected Claude Mythos architecture, built from public research papers with no Anthropic weights.
- The central hypothesis: Mythos is a Recurrent-Depth Transformer that loops the same block of layers up to 64 times per forward pass, reasoning silently in continuous latent space before emitting any token.
- Claude Mythos itself remains locked inside Project Glasswing, where ~50 partners used it to find over 10,000 high-severity software vulnerabilities in critical infrastructure.
- The Parcae paper (UCSD + Together AI) provides the theoretical backbone: a stable looped architecture where a 770M-parameter model matches a 1.3B dense transformer in quality.
- Pre-configured model variants from 1B to 1T parameters are included; install with
pip install open-mythosand a training script on FineWeb-Edu is ready to run. - The repo is explicitly speculative — Anthropic has published no architecture paper, and real Mythos may differ significantly from OpenMythos's reconstruction.
Claude Mythos is Anthropic's most powerful large language model, developed to autonomously find software vulnerabilities. Anthropic has not released it to the public, citing safety and misuse concerns. So when a 22-year-old developer named Kye Gomez dropped a fully runnable PyTorch implementation of what he believes the architecture looks like, the community took notice. Gomez published OpenMythos on GitHub, and the repository accumulated more than 10,000 GitHub stars in roughly two weeks.
The model Anthropic won't show you
Anthropic and its Project Glasswing partners have identified more than 10,000 high- or critical-severity vulnerabilities in critical software systems. Claude Mythos Preview is a large language model that can autonomously find zero-day vulnerabilities and create exploits for them. Anthropic instead launched Project Glasswing, with a consortium of companies using Mythos to find and fix software vulnerabilities. Anthropic is committing up to $100M in usage credits and $4M in donations to open-source security organizations to support this work.
The model's capabilities in specific areas like cybersecurity and biology are advanced enough that they could be misused to create cyberattacks or dangerous weapons. Anthropic has not published any technical paper on Claude Mythos, prompting speculative theories within the research community. That vacuum is exactly what OpenMythos is trying to fill.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
