Someone Created an Open-Source ‘Theoretical Mythos’ to Transform the Dangerous AI of the Anthropic Engineer

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

A developer named Kye Gomez published it OpenMythosan open revision of what he thinks is Claude Mythos looks under the hood. The repo has garnered more than 10,000 GitHub stars in the weeks since its release, and ships with a “readme” file full of equations, quotes, and a disclaimer that it’s not affiliated with Anthropic.

It’s a fantasy. But it’s an organized fantasy, in code.

Here’s a quick refresher on what Mythos is: Legends went public at the end of Marchwhen Anthropic accidentally published a note describing it as the company’s best model to date—a step above Opus. The follow-up, Mythos Preview, turned out to be very good for cybersecurity.

According to Anthropic, Mythos found 271 vulnerabilities in Firefox during Mozilla’s testing. It became the first AI model to complete a 32-step corporate network attack simulation. Anthropic locked it inside Project Glasswing, an official collaboration of about 40 people, including Microsoft, Apple, Amazon, and the NSA.

People never touch it. So Gomez tried to figure out how it works.

The central idea of ​​OpenMythos is that Mythos is a Recurrent-Depth Transformer—also called a looped transformer. Standard colors include hundreds of unique colors. Curved models pick up particles and propel them often through the forward pass.

In other words, it’s the same weight going over and over again. To think deeply, in the secret place continuously, before any sign is released.

Repo says that this may explain two of the Mythos’ most surprising qualities: It creates new problems no other genre can confuse, but its memorization is not the same. It’s a looping finger — a storage design.

OpenMythos cites Parcae, an April 2026 paper from the University of California San Diego and Together AI that solved a long-standing problem of instability in convoluted models – the 770 million Parcae version is equivalent to a 1.3 billion deep switch in positives, with rules for determining how many loops there are. The repo also borrows DeepSeek’s Multi-Latent Attention for memory compression, and a Mixture-of-Experts implementation to handle all domains.

What it doesn’t have is weight, so it’s an unsupervised method.

OpenMythos is a myth. The code defines a range from 1 billion to 1 trillion, but you have to train them yourself—the readme file points to 3 billion training sessions on FineWeb-Edu and Chinchilla’s updated target of 30 billion, which is the kind of computing bill that runs hundreds of thousands of dollars on H100s. No one has done it yet.

So why is it important?

Because it’s the second time in a month that they’ve brought down the wall around Mythos. The first was a lesson from Vidoc Security, which to be made again several of the vulnerabilities found by Mythos using GPT-5.4 and Claude Opus 4.6 within the open source agent. No Glasswing access, and under $30 per scan. Different aspect, same point: The moat surrounding Mythos may be less than the marketing suggests.

OpenMythos is a iteration of Vidoc performing different functions. Vidoc recreated what Mythos released – a self-discovering vulnerability – using existing models. OpenMythos is trying to recreate architecture – the virtual machine that creates the output. One says you don’t need Mythos to find bugs Mythos found. The other says, eventually, you can create something like Mythos yourself.

Anthropic almost never shares Gomez’s opinion publicly, and several options in OpenMythos are transparently blocked—the readme file ensures that it’s obscure enough to let users know that this is just a method. It repeatedly says “probably,” “doubtful,” and “almost certainly.” True Myths cannot be changed at all. Or it could be the one with the details that Gomez hasn’t changed.

What OpenMythos shows is that research articles already have many pieces. Looped transformers, Mixed of Experts, Multi-Latent Attention, Adaptive Computation Time, Parcae fixed stability-no owner. The repo, more than anything else, is what is publicly known about how to build a Mythos version.

The repo is licensed by MIT, and has 2,700 forks already. The tutorial is sitting there, waiting for someone with a GPU team and an idea to confirm it.