Someone Built an Open-Source ‘Theoretical Mythos’ to Reverse-Engineer Anthropic’s Most Dangerous AI

If Anthropic doesn’t want to show you what’s inside the most dangerous AI, someone on GitHub will guess.

A developer named Kye Gomez published Open Mythosan open-source reconstruction of what he thinks Claude Mythos looks like under the hood. The repository amassed over 10,000 GitHub stars within a few weeks of release and comes with an extensive ‘readme’ file full of comparisons, quotes, and a polite disclaimer that it has nothing to do with Anthropic.

It’s speculation. But it’s structured speculation, in code.

Here’s a quick refresher on what Mythos is: Mythos leaked into the public view in late March, when Anthropic accidentally published draft material describing it as the company’s most capable model yet – a level above Opus. The sequel, Mythos Preview, turned out to be inextricably good in the field of cybersecurity.

According to Anthropic, Mythos found 271 vulnerabilities in Firefox during Mozilla testing. It became the first AI model to complete a 32-step attack simulation on corporate networks. Anthropic locked it into Project Glasswing, a vetted coalition of about forty partners, including Microsoft, Apple, Amazon and the NSA.

The public is never allowed to touch it. So Gomez tried to figure out how it works.

The central premise of OpenMythos is that Mythos is a Recurrent-Depth Transformer, also called a loop transformer. Standard models stack hundreds of unique layers. Loop models take a smaller stack and pass it through many times per forward pass.

In other words, they are the same weights undergoing more iterations. Think more deeply, in a continuous latent space, before emitting a token.

The repo claims this would explain Mythos’ two strangest features: it reasons through new problems that no other model can crack, but its raw memorization is uneven. That’s the architectural fingerprint of looping: composition over storage.

OpenMythos cites Parcae, an April 2026 paper from the University of California San Diego and Together AI that solved the long-standing instability problem in loop models: A Parcae model with 770 million parameters is equivalent in quality to a transformer with a fixed depth of 1.3 billion, with predictable scaling laws for the number of loops to execute. The repository also borrows DeepSeek’s Multi-Latent Attention to compress memory, and a Mixture-of-Experts setup to handle the breadth of domains.

What it doesn’t have are weights, so basically it’s a technique without a performer.

OpenMythos is theoretical. The code defines model variants from 1 billion to 1 trillion parameters, but you have to train them yourself. The readme file references a 3 billion parameter training script on FineWeb-Edu and a Chinchilla custom target of 30 billion tokens, the kind of computing account that can run into the hundreds of thousands of dollars on H100s. No one has done it yet.

So why does it matter?

Because it’s the second time in a month that someone breaks the wall around Mythos. The first was a study from Vidoc Security, which reproduced some of Mythos’ most alarming vulnerability findings using GPT-5.4 and Claude Opus 4.6 in an open source agent. No Glasswing access and less than $30 per scan. Different angle, same conclusion: the moat around Mythos may be thinner than the marketing suggested.

OpenMythos and the Vidoc replication do different tasks. Vidoc reproduced the results of Mythos (the discoveries of the vulnerability itself) using existing models. OpenMythos attempts to reproduce the architecture: the actual machine that produces this output. One says you don’t need Mythos to find the bugs Mythos found. The other says that you might eventually be able to build something like Mythos yourself.

Anthropic almost certainly doesn’t share Gomez’s architectural assessments publicly, and several design choices in OpenMythos are explicit fences; the readme file makes sure it’s vague enough so that users know this is just an approximation. It repeatedly says ‘probable’, ‘suspect’ and ‘almost certain’. Real Mythos may not be a loop transformer at all. Or it could be one with details that Gomez hasn’t reverse-engineered yet.

What OpenMythos shows is that the research literature already contains most of the pieces. Loop transformers, expert mix, multi-latent attention, adaptive computation time, Parcae’s stability solution: none of this is proprietary. Above all, the repo is an inventory of what is publicly known about building a Mythos class model.

The repo is MIT licensed and already has 2,700 forks. The training script is there waiting for someone with a GPU cluster and a thesis to provide the proof.