Breaking

Forensic Analysis Says Purported Leaked Anthropic 'Mythos' Model on Hugging Face Appears to Be a Fake

November 2, 2026 at 18:46 EST

Security
Foundation Models

An independent forensic write-up argues that a 7.72TB model checkpoint posted on Hugging Face and presented as a leaked Anthropic 'Mythos' model looks fabricated rather than authentic.

FORENSIC REVIEW · burner22/numbat

A 7.72TB "leaked Mythos" checkpoint looks fabricated, not trained

An outside forensic analysis reports that the file can be rebuilt bit for bit from a SHA-256 hash and NumPy PCG64 random seeds. Output of that kind is not what real training would leave behind.

7.72TB

Size of the posted checkpoint

633

Tensors in the file

11.37M

Values across those tensors

373.6M

Weight signs matched by the reconstruction

Reported counts on one scale

Column height is proportional to each reported figure. The write-up does not fully specify what the 373.6M signs were compared against, so read this as a scale of the reported numbers, not a verified ratio.

373.6M

Matched weight signs

11.37M

Values in the file

How the reconstruction works

STEP 1

Short seed and SHA-256 hash

→

STEP 2

NumPy PCG64 regenerates the weights

→

STEP 3

Output matches the file bit for bit

→

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.