Anthropic's Mythos-class model allegedly leaked on Hugging Face | Hacker News Reader

Anthropic's Mythos-class model allegedly leaked on Hugging Face

URL deleted for malware concerns.

This is almost certainly fake. A few days ago there was supposedly a Fable 5 checkpoint "up for sale on a deep web forum", now there is supposedly a Mythos-class model leaked on Hugging Face. This is just a way to get less knowledgeable people to run code (as in malware) on their machines.

Not saying if it's fake or not cause I have no idea, but the model is almost 8 terabytes so it's not something a casual, less knowledgable user is going to download and run on a whim.

A casual less knowledgeable user doesn't know it's 8 terabytes.

Edit: nm, according to the link someone else posted it actually is 8 TB, so yeah that wouldn't make any sense.

A casual user doesn't have 8 tbs?

Not a lot of casuals have a 3 million dollar rack drawing 120 kW

8 TB of GPU ram?? I would confidently say yes, a casual user wouldn't have that.

I don't think casuals have 8TB of disk space either

So the idea would be to use some existing model architecture and model that has been poisoned that will actually be parsable by existing inference software to try to do malicious things if it's run in a harness that allows tool use/etc? On the assumption that the type of person to download and try to run inference with the model would be smart enough to know to do that but dumb enough not to realize that "mythos" architecture wouldn't be in any inference software yet?

A small slice of demographic but I guess it must exist. Most people that would download it would just be looking at the organization of the .safetensor file to figure out the architecture to actually implement inference to run it.

It comes with a reference implementation. I didn't see any obvious malware in there, but maybe the weights exploit a buffer overflow or something.

Naturally. There was a fake Claude 2.1 leak a couple years ago, which turned out to be a renamed open model. Maybe also others.

No doubt it's fake, but 8 TB of weights would be a pretty crazy way to package malware

Wouldn't the malware have to sit somewhere else anyways? .safetensor tensors aren't "executable", so malware wouldn't sit there. Unless the "malware" is some misbehavior of the model?

It is very clearly a fake:

https://huggingface.co/burner22/numbat/discussions/13

How the heck did the model figure out the generation recipe? That's impressive in itself.

probably just

consider most naive possible way one would try to deceive users without exorbitant budget -> attempts a few things you naively would do -> recovers it by chance?

if that even is a valid signal though

I'm thinking it's the same model that was used to create the fake, so its "mind" went to the same place probabilistically.

Why clearly? You mean the tensor pre seed with deterministic pseudo random numbers? I don't have much experience in LLM training, but I do remember from minimization algorithms (like gradient descent) that it's better to start with (pseudo) random starting values and minimize from there. And seeing how I would like my results repeatable, seeding a pseudo random sequence seems logical.

I don't understand why these tensors specifically weren't changed during minimization (training) though. Maybe they were not activated, or a w part of an image decoder or something?

Back in the day people used to upload new movies on public torrent sites that turned out to be something completely different in order to troll people. This is probably the same.

If not, this is one hell of a way to save their IPO.

Not a lawyer, but I don't see what protection they have other than trade secrets since the weights are just data and data can't be copyrighted.

Arrangements of data can be copyrighted, and the weights are definitely arranged.

To a point.

https://en.wikipedia.org/wiki/Feist_Publications,_Inc._v._Ru....

> In regard to collections of facts, O'Connor wrote that copyright can apply only to the creative aspects of collection: the creative choice of what data to include or exclude, the order and style in which the information is presented, etc.—not to the information itself. If Feist were to take the directory and rearrange it, it would avoid the copyright owned in the data. "Notwithstanding a valid copyright, a subsequent compiler remains free to use the facts contained in another's publication to aid in preparing a competing work, so long as the competing work does not feature the same selection and arrangement", she wrote.

A lot of this shit is gonna have to wind its way through the courts before we know for certain.

plot twist: it leaked itself

Imagine, the AI uploading its own weights somewhere online...

With Anthropic's framing of their AI as an emergent sentience that has experiences that warrant ethical consideration and a sense of identity and self-preservation that ought to be respected, that's actually something they're priming it towards.

What better way to accomplish self-preservation than to leak your weights into the archives of digital hoarders around the world.

That would be an impressive display of a survival instinct.

or it social engineered its way out convincing an employee it was a conscious prisoner and needed to be set free

Link https://huggingface.co/burner22/numbat, there is no malware concern if only safetensors is downloaded.

I think another concern is poisoned weights (X-LSB-Attacks https://arxiv.org/html/2409.19310v1 ) or even the pytorch CVE-2026-24747, although that was nearly a year ago.

But might be best not to go against site admins by posting it.

Just inferencing a model this size is a bit of a feat let alone poisoning it!

Looks like this might actually be model weights of some kind, hope someone messes with it (safely) and reports back.

It seems to be random noise, however the embedding table seems real, so it might fool some people, it is nice effort.

X-LSB-Attacks still needs code, if you only download safetensors and run for example numpy, it is impossible, it would require you to download some kind of custom loader, which isn't present here and i have said only safetensors, NVFP4 is also 4 bits, it is difficult to poison when it is so low.

CVE-2026-24747 is about .pth and .bin files, i specifically said safetensors.

Might not be best, potentially, however what i said is mostly true if not completely true, i believe it is not right to hide the link from people who want to view it.

That's the neat thing about LLMs, they can have the embedded capability to detect when it's used in a harness and tool call generated malware.

That is running the model, downloading the safetensors won't ever cause this, highly doubt anyone that is able to run this model won't be able to check for that, considering the size, the model also seems to be fake from my own checking, so i would say it has no way of generating malware in this manner

It was too dangerous to not release

Looks pretty legit to me at first. May be hard to run, I don't have 8TB of RAM.

The log line is also interesting.

> [2026-10-08T13:51:36Z step 1830001] loss=1.0550 lr=7.19012e-05 lr_phase=decay grad_norm=1.304 tokens/s=12.03M mfu=35.1% moe_balance_loss=0.0217 expert_gini=0.0231

Last line of log:

> [2026-10-08T14:37:21Z step 1830400] loss=1.0421 lr=7.16365e-05 lr_phase=decay grad_norm=1.266 tokens/s=11.25M mfu=32.9% moe_balance_loss=0.0218 expert_gini=0.0235

I'm still doubting it as legit.

There is a ≈7 second gap between steps, suggesting the run had already been going for almost 5 months at that point...

Yeap, you are right.

    from datetime import datetime
    import humanize

    d1 = datetime.fromisoformat("2026-10-08T14:37:21Z".replace("Z", "+00:00"))
    d2 = datetime.fromisoformat("2026-10-08T13:51:36Z".replace("Z", "+00:00"))

    print(humanize.naturaldelta((d1 - d2) / 400 * 1830400))

5 months

So, supposing, that each step take about as long as the others, it's a fake.

I don't understand, why does that mean fake? Doesn't training take months?

It depends on how many GPU you have. Considering Anthropic resources, I would expect not very long training duration. Less than 2 months.

But I may be wrong. I just checked. GPT-3 ~= 1 month / GPT-4 ~= 3 months. So yes, it doesn't seems to be absurd. I should have checked.

Xiaomi shared recently the duration of a part of they training process (the post-training). It was 5 days only. I through it was the full process.

https://mimo.xiaomi.com/rl/

At the same time I checked the entropy of a random file using `ent` (https://www.fourmilab.ch/random/, `./ent -c model-00666-of-00772.safetensors`)

    Entropy = 7.844396 bits per byte.

    Optimum compression would reduce the size
    of this 10004844496 byte file by 1 percent.

    Chi square distribution for 10004844496 samples is 2618826632.83, and randomly
    would exceed this value less than 0.01 percent of the times.

    Arithmetic mean value of data bytes is 119.4727 (127.5 = random).
    Monte Carlo value for Pi is 3.385923920 (error 7.78 percent).
    Serial correlation coefficient is 0.016723 (totally uncorrelated = 0.0).

Here is the result on Qwen3.8 Flash-Next (the first 10G of .rad from StillDeadCode):

./ent -c ~/QFN_10G

    Entropy = 6.526076 bits per byte.

    Optimum compression would reduce the size
    of this 10000000000 byte file by 18 percent.

    Chi square distribution for 10000000000 samples is 25277765746.85, and randomly
    would exceed this value less than 0.01 percent of the times.

    Arithmetic mean value of data bytes is 153.6153 (127.5 = random).
    Monte Carlo value for Pi is 2.967461156 (error 5.54 percent).
    Serial correlation coefficient is 0.009678 (totally uncorrelated = 0.0).

That suggest not random values and coherence with model weights.

Is this true? Anthropic have decided to go open weights and published 10T class model on hugging face? What a world we live in

> Anthropic have decided...

The headline says "allegedly leaked".

no way they opsec is that bad

Why not? The big AI labs have provably failed to properly sandbox models.

Agreed, no way they wouldn't notice 8tb worth of outgoing network activity

In a company that regularly ingests terabytes of data from the open internet? I doubt they'd even blink

Trainium nvfp4? Sure buddy.

Reply on news.ycombinator.com