Fake Claude leak: 43 data blocks were random

A Hugging Face page that claimed to hold a leaked Anthropic AI model is almost certainly fake, users found on the 1st of November 2026. Hugging Face is a website where people share AI models.

A user showed that 43 blocks of the model's data were plain random numbers. A real trained model would not match them.

What the page claimed

The page, called burner22/numbat, said it held an unfinished copy of Anthropic's next Mythos model, explainx.ai reported. Its files add up to about 7.72 terabytes.

The page had about 5,800 downloads when explainx.ai checked it. Anthropic has not commented.

How users checked

A user called FloxUnder showed that 43 data blocks could be rebuilt from their names with a simple random number tool. All 730,368 numbers matched exactly.

Other users said they got the same result. Only small parts of the files were downloaded for the test.

A warning about code

The page includes its own computer code to load the model. Running code from an unknown page can let it take over your computer, explainx.ai warned.

explainx.ai advised people not to download or run it.

Below we break down the test

The 43 blocks were 28 of one kind and 15 of another, the audit shows. Every one of the 730,368 numbers matched, a 100 percent match by our arithmetic.

What is confirmed and what is not

The audit did not check every block, its author said. Nothing on the page links it to Anthropic.

Key facts

A Hugging Face page claimed to hold a leaked Anthropic Mythos model.

Users found 43 data blocks were made of random numbers.

Anthropic has not commented.

explainx.ai advises people not to download or run the files.