- Developer Kye Gomez has published OpenMythos, an open-source architectural guess at Anthropic‘s unreleased Claude Mythos AI model.
- The model was locked away in Project Glasswing after autonomously finding 271 Firefox vulnerabilities and completing a 32-step network attack.
- The code is a theoretical reconstruction using public research, lacking trained weights, and would cost hundreds of thousands of dollars to train.
- This follows a separate effort by Vidoc Security that reproduced Mythos’s vulnerability findings using existing models for under $30 per scan.
A developer has published a speculative, open-source blueprint for one of the world’s most restricted AI models, according to a GitHub repository. Kye Gomez’s OpenMythos project attempts to reconstruct the architecture of Anthropic‘s powerful Claude Mythos, which the company locked away after its alarming Cybersecurity capabilities were revealed in March.
The model, described in leaked materials as more capable than Opus, was confined to Project Glasswing, a vetted coalition including Microsoft, Apple, and the NSA. Consequently, the public has no access, prompting this crowdsourced effort to understand its design. Gomez’s central guess is that Mythos is a Recurrent-Depth Transformer, a looped architecture enabling deeper reasoning before emitting any token.
This theoretical scaffolding cites a 2026 paper called Parcae for stability and borrows techniques like DeepSeek‘s Multi-Latent Attention. However, the repository contains only code and no trained model weights, making it an executor without a technique. Training a version would reportedly cost hundreds of thousands of dollars, a barrier that so far remains unbreached.
Meanwhile, a separate study from Vidoc Security reproduced several of Mythos’s vulnerability discoveries using GPT-5.4 and Claude Opus. This replication, conducted for under $30 per scan, suggests the security moat around Mythos may be thinner than suggested. The two projects tackle different challenges: one replicates outputs, while the other attempts to reverse-engineer the machine itself.
The OpenMythos readme is filled with careful disclaimers, using words like “likely” and “suspected.” It ultimately serves as an inventory of publicly available research that could build a Mythos-class model. With an MIT license and thousands of forks, the project now waits for someone with sufficient compute and curiosity to bring the theory to life.
✅ Follow BITNEWSBOT on Telegram, Facebook, LinkedIn, X.com, and Google News for instant updates.
