OpenMythos: Open‑Source Reverse‑Engineering of Claude Mythos Architecture and the Controversy

1. Project Overview

OpenMythos is an open‑source project started by independent developer Kye Gomez, claiming a theoretical reconstruction of Anthropic’s Claude Mythos based on publicly released papers. It is implemented in pure PyTorch, installable via pip, and provides pre‑configured models ranging from 1 B to 1 T parameters with up to 64 inference loops.

2. Core Architecture: Recurrent‑Depth Transformer

According to the OpenMythos documentation, Claude Mythos uses a Recurrent‑Depth Transformer (RDT), which differs from traditional deep‑stacked Transformers by re‑using the same weight set multiple times within a single forward pass.

2.1 Three‑Stage Inference

Prelude : standard stacked Transformer blocks for initial encoding.

Recurrent Loop : core innovation; selected layers are executed up to 64 times without emitting intermediate tokens, reducing latency.

Coda : final layer that decodes the hidden state into output tokens.

2.2 Attention and Sparse MoE

The implementation can switch between MLA (Multi‑Head Latent Attention) from DeepSeek, which compresses KV caches via low‑rank factorisation, and GQA (Grouped Query Attention) from Google’s Llama. The feed‑forward network uses a sparse mixture‑of‑experts (MoE) with routing and shared experts.

2.3 Parameter Efficiency

OpenMythos reports that the 770 M‑parameter model achieves performance comparable to a conventional 1.3 B‑parameter model, a reduction of roughly 40 % in parameters.

3. Community Reception

The release sparked polarized reactions. Supporters view it as a strong open‑source response to closed‑source AI labs, arguing that systematic analysis of public papers can reconstruct core techniques and benefit AI education and research. One developer praised the effort as “no black box can stop people from tearing it apart”.

Critics point to the developer’s history of repackaging others’ work, noting inflated star counts and questionable code quality. Some compare the narrative to previous controversies, accusing the project of over‑hyped storytelling. Others stress that the work is a “theoretical reconstruction” rather than true reverse engineering, emphasizing the conceptual difference.

4. Blue‑Team Perspective

From a defensive security standpoint, the engineering of recurrent inference lowers the barrier for reproducing deep‑thinking AI models; a 7 B‑level RDT model can run on a single consumer‑grade GPU, facilitating research on AI inference behavior and enabling more precise anomaly‑detection baselines.

For AI‑driven vulnerability‑scanning operators, an open‑source Claude‑Mythos‑level model could accelerate automated zero‑day discovery, raising the stakes for vulnerability‑management processes and patch‑response SLAs. However, the claim that Claude Mythos can autonomously discover zero‑day bugs remains loosely defined, and the actual capabilities within Project Glasswing are undisclosed.

5. Conclusion

OpenMythos demonstrates the feasibility of implementing a Recurrent‑Depth Transformer and provides a reference artifact for researchers studying frontier AI architectures. Nevertheless, the project’s marketing narrative blurs the line between “theoretical reconstruction” and “black‑box reverse engineering”. For security teams, the lowering of replication costs signals an emerging AI‑assisted vulnerability‑exploitation arms race that requires updated defensive strategies.

References: OpenMythos GitHub repository (github.com/kyegomez/OpenMythos); X post by @IndieDevHailey, 2026‑06‑28.