A Hugging Face account named burner22 posted a giant file late this week. The account called it numbat, Anthropic's next model. Readers on Hacker News and Reddit tested that claim fast1.
By evening, the math had answered them. The file was fake. Someone built it, piece by piece, to pass a quick glance as a real leak, counting on curiosity to outrun scrutiny12.
A Fake Anthropic Leak Pulled 5,846 Downloads Before Readers Caught It
Counts of downloads, likes and discussion signals on burner22's Hugging Face page and the Hacker News thread that debunked it, Nov. 1, 2026
Each bar is one count tied to the file: the magenta bars show how far it spread on Hugging Face, and the gray bars show the Hacker News thread that caught it.
Sources: Hugging Face [1]; Hacker News [2]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Hugging Face downloads | 5,846 |
| Hugging Face likes | 105 |
| Hacker News points | 32 |
| Hacker News comments | 13 |
A File That Looked Like Anthropic's Next Model
The upload held 772 shards of data, each one under 10.6 gigabytes. Together they claimed 7.72 terabytes, enough for roughly 14 trillion parameters, the settings a model learns during training1.
burner22 backed the claim with more than a name, listing a loader, a tokenizer, and a checkpoint step. Each piece aimed to make the file look ready to run. It even offered its own proof. "Every tensor in every shard is internally consistent: E4M3 scale blocks decode their E2M1 payloads within RTNE bounds, block scales never underflow, and every NVFP4 quad round-trips exactly," the card read1.
Two questions in the page's FAQ gave away the uploader's mood. Asked how they got the file, burner22 wrote, "Can't say more than what's already in the repo." Asked when the rest would land, the answer came back short: "They're all uploading. Be patient"1.
The codename did real work for the claim. Numbat turned up in March, inside Claude Code's leaked source files, naming an unreleased Anthropic project still in testing3. That leak was real. Anthropic later called it "a packaging issue caused by human error," after roughly 500,000 lines of code reached the public npm registry4.
The Math That Gave It Away
Numbers broke the claim first. Hugging Face's own scan ignores whatever an uploader claims. It put the file at 6.85 trillion parameters, well under burner22's listed 14 trillion1.
A Hacker News reader named kingstnap opened the shards directly and checked the numbers inside them by hand. "Its fake, the matrixes are literally just random data and the seed is the tensor name," kingstnap wrote2. The math settled it.
burner22 Claimed 14 Trillion Parameters. Hugging Face Counted 6.85.
Parameters inside the numbat upload, the uploader's claim against Hugging Face's own tensor count, Nov. 1, 2026
The blue dot is the uploader's claimed parameter count, the magenta dot is what Hugging Face's own scan found inside the file, and the rod spans the gap between them.
Source: Hugging Face [1]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Uploader's claim | Hugging Face's count |
|---|---|---|
| Parameters in the numbat file | 14.00 trillion | 6.85 trillion |
Other commenters found the same seam from different angles. One matched 43 whole tensors to plain NumPy random-number output, bit for bit. The match points to a generator, past any training run.
A second reader caught a label fight. The file called itself NVFP4, a four-bit format Nvidia built for recent chips, but Hugging Face's own scan tagged the same file "8-bit" instead2. Two labels disagreed on one page.
explosion-s, a third commenter, added the detail that stings most. "Anthropic DMCA'd the Claude Code mirrors within days. This is still up," they wrote, setting Anthropic's fast response to its real leak against its silence on this one2. The gap said plenty.
A Real Codename Dressed Up a Fake File
Here is the strongest case for burner22's claim. numbat is a genuine Anthropic project name, confirmed by a leak Anthropic itself has left unchallenged since March34. A hoax inventing its own codename would have drawn suspicion inside minutes. This one skipped that step.
That strength is also the tell. Anyone reading March's coverage could copy a real name into a fresh upload, well before burner22 found it. The leak was public and widely covered.
Matching Hugging Face's own parameter count takes an actual trained model, the step a forger skips every time. Training at that scale costs real money and real months. The count, past the name, is what settled this for every reader who checked it1.
The same day carried a smaller version of the same question. A screenshot on X showed a usage line reading "gpt-6.2-astra" inside OpenAI's Codex billing data. Within the hour, a second account posted a matching copy67.
Token Gremlin@TokenGremlinOpenAI is quietly A/B testing a new variant called GPT-6.2 Astra, and it's already showing up in Codex usage analytics.
The model picker still says GPT-6 Astra, but the usage breakdown for some accounts now lists gpt-6.2-astra separately. Multiple users are seeing the exact same… pic.twitter.com/XAmFfL27m6
OpenAI stayed silent about the name, and its public model list still skips it entirely6. A screenshot proves a picture exists. Proving what the picture shows takes independent checking, the same gap that sank numbat. Every reader of a "leak" owes the story that same check before sharing it.
What to do
Check a platform's own numbers against a claim before repeating it. Hugging Face's scan sat on the same page as burner22's upload the whole time, free for anyone to read. Checking it took one look, before any download.
Treat a familiar codename as a reason to look closer, instead of a reason to relax. A detail a forger copied from a real leak proves the forger did homework, past proving the file is real. Copying a name takes minutes. The tensor math behind it takes real training to produce.
Wait for a second named check before sharing a "leak" further. kingstnap's read took minutes once someone opened the shards directly. The same fast check catches most forged files circulating online.
By the numbers
- Hugging Face's own scan counted 6.85 trillion parameters inside burner22's file1.
- burner22's model card claimed 14 trillion parameters, almost double the scan1.
- The upload split its claimed weights across 772 shards, each under 10.6 gigabytes1.
- Those shards added up to 7.72 terabytes of claimed data1.
- Downloads on burner22's page reached 5,846 before the debunking caught up1.
- Likes on the same page stood at 105 as of this run1.
- The Hacker News thread debunking the file earned 32 points and 13 comments2.
- One reader matched 43 tensors to plain random-number output, bit for bit2.
What to watch
Anthropic's silence is itself worth tracking. The company moved fast against March's genuine leak. A comparable response here would say whether this fake crossed some threshold inside the company.
The next fake's lifespan matters more than this one's did. This file fell in hours. Its own numbers sat in public, on the same page as the claim, open to anyone who thought to check first.
Watch, as well, whether burner22's account stays up or disappears quietly once the attention moves on. Either path tells readers something about who built this and why.
Sources
- "burner22/numbat," Hugging Face, Oct. 23, 2026, https://huggingface.co/burner22/numbat
- "Alleged Mythos Class Anthropic Checkpoint Leak," Hacker News, Nov. 1, 2026, https://news.ycombinator.com/item?id=50327240
- Boulanouar Walid, "Claude Code Source Leak: What Was Inside," AYAutomate, Apr. 2, 2026, https://www.ayautomate.com/blog/claude-code-source-code-leaked-github-2026
- Beatrice Nolan, "Anthropic source code, Claude Code data leak, second security lapse days after accidentally revealing Mythos," Fortune, Mar. 31, 2026, https://fortune.com/2026/03/31/anthropic-source-code-claude-code-data-leak-second-security-lapse-days-after-accidentally-revealing-mythos
- Dillip Chowdary, "Alleged Mythos Class Anthropic Checkpoint Leak," TechBytes, Nov. 1, 2026, https://techbytes.app/posts/alleged-mythos-class-anthropic-checkpoint-leak/
- Gennaro Cuofano, "GPT-6.2 Astra Shows Up in Codex Usage Screenshots on X," FourWeekMBA, Nov. 1, 2026, https://fourweekmba.com/ai-gpt-6-2-astra-shows-up-in-codex-usage-screenshots-on-x/
- Token Gremlin (@TokenGremlin), post on X, Nov. 1, 2026, https://x.com/TokenGremlin/status/2116834891744600291
