Developer Open-Sources Reconstruction of Anthropic's Mythos Architecture

Developer Kye Gomez published the open-source project OpenMythos, which the repo and multiple news outlets describe as a public reconstruction that approximates Anthropic's Claude Mythos (Forbes, Yahoo Tech, Decrypt). Yahoo Tech reports the repository exceeded 10,000 GitHub stars within weeks and includes an extensive readme, equations, and runnable code (Yahoo Tech). Forbes and Yahoo Tech report that Anthropic did not release Claude Mythos publicly and instead provided the model to a vetted coalition called Project Glasswing of about 40 partners; Yahoo Tech reports that Anthropic said Claude Mythos found 271 vulnerabilities in Firefox during Mozilla testing (Yahoo Tech).

The OpenMythos repo frames the central hypothesis as a Recurrent-Depth Transformer (RDT), also described in coverage as a looped transformer where a smaller weight stack is iterated multiple times per forward pass rather than stacking many distinct layers (Decrypt, Yahoo Tech, 36Kr). The reconstruction combines three architectural elements attributed in reporting: looped depth, a Mixture-of-Experts (MoE) routing mechanism to provide breadth, and memory-compression or multi-latent attention ideas drawn from recent research including the Parcae paper and DeepSeek-style modules (Decrypt, 36Kr). Decrypt and 36Kr note that Parcae (a UC San Diego and Together AI paper) reportedly addressed prior instability in looped models and demonstrated favorable scaling laws, which the repo cites as an enabling component.

Industry observers note that stacking loops rather than parameters is a recurring design motif in recent research on efficient depth. Comparable public research shows looped or recurrent-depth approaches can match fixed-depth models with fewer parameters when combined with stabilization techniques and conditional computation such as MoE. For practitioners, the practical trade-off is between repeated compute per token and reduced parameter count, which alters deployment cost profiles and hardware utilization patterns in ways different from parameter-scaling strategies.

Reporting frames OpenMythos as significant because it synthesizes multiple public papers and community speculation into runnable code quickly after Claude Mythos became headline news. Observers quoted in coverage highlight two broader implications: first, that frontier architectural ideas can be reassembled from public research and limited disclosures; second, that capability diffusion may outpace centralized control when motivated developers and accessible compute combine. These are general observations about the ecosystem and are not claims about Anthropic's internal choices.

Practitioners and security teams should watch for follow-up community experiments that benchmark OpenMythos variants against standard baselines, replication studies of the Parcae stability claims, and incidents where reassembled models are applied to cybersecurity tasks. Reporting indicates the repo's claims and design choices are speculative reconstructions and that OpenMythos is explicit about being independent from Anthropic (Yahoo Tech, Decrypt). Observers will also track whether academic and industry groups reproduce the looped-model scaling behavior and whether open-source implementations surface novel misuse concerns.

Organizations focusing on model risk and red-teaming will treat runnable reconstructions plus public research as new inputs to threat models. The pattern in coverage is that architecture-level innovations migrate quickly when papers and community implementations align, changing the set of artifacts defenders must monitor.