Viral "Anthropic Mythos leak" on Hugging Face holds random numbers, not trained weights

A 7.72 TB repository on Hugging Face that spread on X on November 1 as a leaked mid training checkpoint of Anthropic's next Mythos class model contains weights generated by a random number generator, not by training, according to a reproducible audit posted in the repository's discussions.

The audit found that 43 complete tensors, 730,368 BF16 values, can be regenerated bit for bit from their tensor names alone. The recipe hashes each tensor name with SHA-256, uses the first eight bytes to seed NumPy's PCG64 generator, draws normal values and applies a fixed scale. The auditor also recorded predictions for 21 tensors that had not yet been read, and every predicted value matched. Other users in the discussion said they could reproduce the result.

CellCog, which sells AI agent services running on Claude, said it reran the recipe and matched all nine tensors it checked in the first shard, 147,968 values, while a wrong seed matched none.

The repository, burner22/numbat, describes a model of about 14 trillion parameters and 114 layers and calls numbat an Anthropic internal codename. It was created on October 23 by an account named burner22, and had 5,846 downloads by the evening of November 1, CellCog reported. Anthropic had not commented.