Fast Facts
AI model leak verification is the habit of testing a claimed leak before reacting to it. A Hugging Face repo called burner22/numbat claims to hold a 14-trillion-parameter checkpoint from “Anthropic’s next Mythos-class training run,” at about 7.72 TB. An audit dated November 1 found 43 of its tensors match a seeded random-number recipe exactly.
That finding does not prove the repo is fake, and nothing links it to Anthropic. It does show how a small, safe test can change the story. Use the seven checks below on the next claimed leak.
Every few months a repository appears claiming to be a frontier model’s secret weights. Reactions run from panic to trading. Careful AI model leak verification costs minutes and avoids both. This guide uses the numbat case as a worked example, with its limits stated.
The Numbat Claim in Brief
The model card describes about 14 trillion parameters, 114 layers, 1,224 experts in 4-bit format and 772 shards. Hugging Face’s own counter shows 6.85 trillion because 4-bit values are packed into bytes, and by FourWeekMBA’s arithmetic that would unpack to about 13.67 trillion, close to the claim. The repository was created on October 23, by an account opened on October 16, and showed around 5,800 downloads.
43 tensors | 730,368 values | about 1 in 9.4 million
An auditor regenerated these values exactly from a name-seeded random generator. The “1 in 9.4 million” share of parameters is FourWeekMBA’s arithmetic. It is a sample, not a full scan.
Seven Checks for AI Model Leak Verification
- Read only the header. A safetensors file begins with a small JSON header, which a ranged HTTP request can fetch without downloading terabytes.
- Sample small tensors. Normalization weights and tiny kernels are cheap to inspect.
- Try a seeded recipe. The audit hashed each tensor’s name to seed a generator and compared bytes. Exact matches suggest generation rather than training.
- Inspect distributions. Trained weights show heavy tails and layer-to-layer drift. Perfect Gaussians with constant variance are a red flag.
- Check the uploader. Account age and history matter.
- Look for official word. No Anthropic comment has been found, and no takedown was reported as of November 1.
- Never run it unsafe. Avoid trust_remote_code and use an isolated sandbox with no credentials.
What the Audit Cannot Show
The audit sampled 43 tensors out of a repository of billions of values. A training setup that deliberately froze some parameters could produce the same pattern, and one commenter reported that some tensors do not match the recipe. The codename “numbat” also appears in publicly circulated Claude Code source, per the explainx write-up, so the name proves little about access. Some outlets already call the leak fake; the evidence supports a narrower claim, which is that the checked tensors look generated. Good AI model leak verification keeps that distinction.
Why AI Model Leak Verification Matters Beyond the Gossip
Fake leaks are bait: they can carry malware, scam buyers or move markets. They also muddy real incidents. See our analysis where we explain AI model weight security and six questions to ask any AI vendor, since real weights are guarded very differently. Gray markets exploit the same appetite: see why discounted AI API access is a data leak with a coupon code. For tampered training data, see an AI data poisoning attack launched against a model.
Access to powerful models is also tiered, which feeds the rumor cycle. See our analysis where we explain AI cybersecurity access tiers and who gets the best AI, and the five costly AI agent security blind spots no one is fixing.
⚠️ Hypothetical scenario: A Lagos fintech’s security lead sees a viral post about leaked frontier weights and an engineer proposes downloading them to “test.” The lead blocks it, runs the header and small-tensor checks from a sandbox, and records the result. Disciplined AI model leak verification avoided a malware risk and a wasted week.
💡 CreedTec Analyst’s Note by Daniel Ikechukwu
Strategic Impact
AI model leak verification is becoming a skill every AI buyer and security team needs. Claims are cheap; cheap tests are available, and they work before the vendor responds.
Stop / Start / Watch
- Stop: forwarding or trading on a leak claim before any check.
- Start: keeping a one-page checklist for claimed leaks, with a sandbox rule.
- Watch: whether Hugging Face acts on the repo and whether Anthropic comments.
ROI Outlook
The cost is minutes of analyst time; the avoided cost is a bad decision or a compromised machine. We cannot put a general figure on it.
— Daniel Ikechukwu
