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

If Anthropic will not present you what is inside its most harmful AI, any individual on GitHub will guess.

A developer named Kye Gomez has printed OpenMythos, an open-source reconstruction of what he thinks Claude Mythos seems like below the hood. The repo has picked up over 10,000 GitHub stars in a number of weeks upon launch, and ships with an exhaustive “readme” file stuffed with equations, citations, and a well mannered disclaimer that it has nothing to do with Anthropic.

It is hypothesis. Nevertheless it’s structured hypothesis, in code.

Right here’s a fast refresher on what Mythos is: Mythos leaked into public view in late March, when Anthropic by accident printed draft supplies describing it as the corporate’s most succesful mannequin to this point—a tier above Opus. The follow-up, Mythos Preview, turned out to be unreleasably good at cybersecurity.



Per Anthropic, Mythos discovered 271 vulnerabilities in Firefox throughout Mozilla testing. It grew to become the primary AI mannequin to finish a 32-step company community assault simulation. Anthropic locked it inside Undertaking Glasswing, a vetted coalition of about 40 companions, together with Microsoft, Apple, Amazon, and the NSA.

The general public by no means will get to the touch it. So Gomez tried to determine the way it works.

OpenMythos’s central guess is that Mythos is a Recurrent-Depth Transformer—additionally referred to as a looped transformer. Customary fashions stack a whole bunch of distinctive layers. Looped fashions take a smaller stack and run it by way of itself many instances per ahead cross.

In different phrases, it’s the identical weights going by way of extra iterations. Deeper pondering, in steady latent area, earlier than any token will get emitted.

The repo argues this might clarify Mythos’s two strangest qualities: It causes by way of novel issues no different mannequin can crack, however its uncooked memorization is uneven. That is the architectural fingerprint of looping—composition over storage.

OpenMythos cites Parcae, an April 2026 paper from College of California San Diego and Collectively AI that solved the long-standing instability drawback in looped fashions—a 770 million-parameter Parcae mannequin matches a 1.3 billion fixed-depth transformer on high quality, with predictable scaling legal guidelines for what number of loops to run. The repo additionally borrows DeepSeek’s Multi-Latent Consideration to compress reminiscence, and a Combination-of-Consultants setup to deal with breadth throughout domains.

What it doesn’t have is weights, so principally it’s a way with out an executor.

OpenMythos is theoretical. The code defines mannequin variants from 1 billion to 1 trillion parameters, however it’s important to prepare them your self—the readme file factors to a 3 billion parameter coaching script on FineWeb-Edu and a Chinchilla-adjusted 30 billion-token goal, which is the sort of compute invoice that runs into a whole bunch of hundreds of {dollars} on H100s. No person’s carried out it but.

So why does it matter?

As a result of it is the second time in a month any individual has chipped on the wall round Mythos. The primary was a research from Vidoc Safety, which reproduced a number of of Mythos’s most alarming vulnerability findings utilizing GPT-5.4 and Claude Opus 4.6 inside an open-source agent. No Glasswing entry, and at below $30 per scan. Totally different angle, identical conclusion: The moat round Mythos could also be thinner than the advertising and marketing urged.

OpenMythos and the Vidoc replication are doing totally different jobs. Vidoc reproduced Mythos’s outputs—the vulnerability discoveries themselves—utilizing current fashions. OpenMythos is attempting to breed the structure—the precise machine that produces these outputs. One says you do not want Mythos to search out the bugs Mythos discovered. The opposite says, ultimately, you may have the ability to construct one thing like Mythos your self.

Anthropic nearly definitely does not share Gomez’s architectural guesses publicly, and a number of other of the design selections in OpenMythos are express hedges—the readme file makes positive to be obscure sufficient so customers know that is simply an strategy. It repeatedly says “doubtless,” “suspected,” and “nearly definitely.” Actual Mythos might not be a looped transformer in any respect. Or it may be one with particulars Gomez hasn’t reverse-engineered but.

What OpenMythos demonstrates is that the analysis literature already incorporates many of the items. Looped transformers, Combination of Consultants, Multi-Latent Consideration, Adaptive Computation Time, Parcae’s stability repair—none of it’s proprietary. The repo is, greater than something, a listing of what is publicly recognized about how you can construct a Mythos-class mannequin.

The repo is licensed MIT, and it has 2,700 forks already. The coaching script is sitting there, ready for somebody with a GPU cluster and a thesis to show.