Anthropic's 14-trillion-parameter Mythos leak turns out to be seeded noise, not trained weights

The leak that wasn't

On November 1, 2026, the AI research community lit up with reports of a massive Hugging Face repository. The repo, burner22/numbat, contained 772 safetensors shards totaling 7.72 TB. Its model card claimed to be a mid-training checkpoint of Anthropic's unreleased Mythos-class model — the successor to Claude Mythos 5, itself pulled from public access in June 2026 after a government order.

The repository was hosted by burner22, a newly created account. Within hours, it had accumulated roughly 5,800 downloads. The claim spread fast: a 14-trillion-parameter frontier model checkpoint, leaked onto a public repository.

But the AI community's response was just as fast. Within hours, independent researchers at CellCog and elsewhere published forensic analyses showing the tensors contained no trained weights. The data matched what you would get from a random number generator seeded with the tensor name itself.

Anthropic had not commented on the repository by deadline. Its news page listed nothing about it.

How the forensics proved it was fake

CellCog's analysis was the most thorough. Their approach: calibrate the random noise pattern against a known reference model (Qwen2.5-32B at step 10587), then test the burner22 shards by seeding a Philox random number generator with bytes drawn from each tensor name.

The match was exact. Each shard's data could be reproduced deterministically from its tensor name alone, using the same noise pattern a seeded RNG would produce. There was no training loss reduction — the tensors had never passed through a training step. The repository was structured noise designed to look like a real checkpoint, complete with convincing metadata, shard sizes, and a polished model card.

Hugging Face discussion #13 posted the name-seeded RNG audit, and commenters linked to CellCog's write-up. By the time the story reached mainstream coverage, the forensic evidence was already public.

The broader implication: tooling for detecting fake AI model checkpoints is now fast and widely distributed. A 14-trillion-parameter, 7.72 TB file set can be debunked in hours — faster than the hype can spread.

The Mythos 5.5 staging leak — a separate incident

Separately on November 1, a reference to 'Claude Mythos 5.5' briefly appeared on Anthropic's official Mythos model page. TestingCatalog documented the mention, and archive checks showed it was not present before that day. The reference was apparently a staging or publishing error — a hidden label that became visible briefly before being removed.

This is a separate incident from the Hugging Face repository. The Mythos 5.5 label is not evidence the burner22 checkpoint was real. If anything, the timing suggests the prankster behind the Hugging Face repository timed the release to coincide with genuine Anthropic activity.

Anthropic has not confirmed the existence or timeline of a Mythos 5.5 model. The March 2026 Fortune article that first reported Mythos described it as a 'step change in capabilities' over Claude Opus, and the June 2026 government order that forced its partial rollback involved both Mythos 5 and Fable 5.

What this means for model security and community trust

The burner22/numbat incident is a case study in how quickly AI hype can outrun reality — and how quickly the community can catch up. A 7.72 TB repository with proper formatting, a convincing model card, and 5,800 downloads was identified as fake within hours because the forensic tooling exists and researchers were willing to run it.

For model security teams, the incident highlights that model weight verification — cryptographic signing of published checkpoints, or at minimum provenance attestation — remains unsolved across the industry. Any large language model provider can have a fake checkpoint uploaded in their name. The current defense depends on community vigilance and rapid forensics rather than systemic protection.

The repository remains on Hugging Face. Its 5,800 downloads mean thousands of people have a 7.72 TB file set that, regardless of its origin, contains no trained model weights.