[gnfargbl]

The most important pieces of information in this report are (a) the confirmation that the PRC models have no guardrails and will participate in offensive activity, and (b) the confirmation that they sometimes meet their objectives.

For the purposes of model selection, it's irrelevant to an attacker if a model achieves an offensive objective 70% of the time, when that model refuses to participate 100% of the time. However, a model that always participates but "only" succeeds 10% of the time is golden -- just run it more often, or give it more tokens. Attackers are patient, and many of them are well-resourced.

[subscribed]

But also the American models (case in point: Fable) refuse to engage in *defensive* activity, so anyone who's not the American government or one of the handful American companies has no choice but to turn to Chinese models to defend themselves.

Not everyone is an attacker, but now the public discourse is "but the evil Chinese will break everything" - yeah, that's because no one is permitted to do vulnerability checks of their own software or infrastructure with the capable models.

Security team in my company is salivating seeing the news, because we have a fighting chance to find and patch many vulnerabilities we didn't previously notice, thanks to the Chinese models.

[fidotron]

You have to wonder if any of these models, from any providers or countries, are set up to lie. i.e. tell you no vulnerabilities while quietly siphoning off the ones they do find into a database.

Yet another reason that self hosted will prove to be the only sane way forward, and it'a almost certainly necessary to have multiple different model providers working adversarially.

[throw10920]

Some PRC models are backdoored to silently insert extra vulnerabilities when certain conditions are met - https://www.boozallen.com/expertise/cybersecurity/whats-in-a...

And that's just the model behavior. The provider itself can do whatever. Given the PRC's public record of prolific IP theft, the default assumption is that they're taking everything you send to one of their APIs.

Anyone have suggestions for poisoning their data?

[Footprint0521]

Booz Allen is cute, but if china can train K3 on a fraction of the US compute availability, yet it benchmarks almost equivalent to Fable for a third of the cost, it’s game over for US labs in the long run.

Believe me, I wish this wasn’t the case, but open up the hardware on the device you are reading this on and tell me the majority of tech inside wasn’t made in China….

Manufacturing was lost a long time ago, this is really just another way

[throw10920]

> if china can train K3 on a fraction of the US compute availability, yet it benchmarks almost equivalent to Fable for a third of the cost, it’s game over for US labs in the long run.

I agree. However, as of yet, most/all leading PRC models are distilled from US models. I've personally observed Deepseek, GLM, and Kimi all respond that they are Claude when asked, and the networks of tens of thousands of proxy accounts that we've found show that it's happening on a large scale.

But - if the PRC labs actually train up the domain expertise to train those models from scratch, which they are in the process of doing - then the US is cooked. They're not there yet, but it's probably only a matter of years.

[Footprint0521]

Playing devils advocate, does from scratch really matter if all frontier labs are training off each other anyways? Practically speaking, businesses/consumers just care about lowest inference cost for maximizing intelligence anyways (not to mention Anthropic/OpenAI forcing KYC/litigation barrier trash for access to any cybersecurity/IT capabilities) that I literally just cancelled Claude today.

Yeah it’s sad the CCP has my information. But it’s either them or the feds, and at least Chinese models actually work for cybersecurity tasks, not to mention aren’t stupid expensive

Artificialanalysis.ai rn on opus 5 is a joke. The main intelligence benchmarks it is like 1% better than Fable, but the cost difference between that and K3 is so funny lol. It’s the same with cars— you don’t have to do it from scratch, as long as you can do it cheaper and with the same quality, hence Toyota/Honda taking over

Yeah it’s sad that American models are censored more than Chinese ones lol (outside of asking them about the CCP) but it’s where we are at I guess :(

[gnfargbl]

Assuming the providers are compromised (and I agree that some of them probably are) then I doubt the angle taken will be to poison the product. That kind of thing usually gets noticed eventually.

A more likely scenario is to focus on the model users as potential victims, e.g. by logging internal infrastructure descriptions, capturing private access tokens from chats, etc. That is very deniable, because it's hard to prove where the compromised data originated.

[subscribed]

Of course, I'm not claiming PRC is a friend of the world. And I agree with your last point, however I don't think it's feasible to self-host Kimi-3 sized inference.

[matheusmoreira]

Self-hosted local models can't become viable soon enough... Currently they require hundreds of thousands of dollars if not millions in capital. That needs to change!

[kouteiheika]

> The most important pieces of information in this report are (a) the confirmation that the PRC models have no guardrails and will participate in offensive activity

This isn't really important for open-weight models, because the guardrails are trivial to remove when you have the weights.

[flyinglizard]

I'm curious about the process. How such a thing (or any kind of model lobotomy) is done?

[roenxi]

It isn't quite a lobotomy, I'm going to butcher the research a bit, but basically what they're finding is that these models create a sort of "bad stuff that I should refuse to engage with" axis in the statistical vector space they operate in. So usually there to be some sort of vector that ranges from 0 for a puppy snoozing peacefully and 1 for writing a virus that exterminates humanity pornographically while broadcasting racist and homophobic slurs (which would be quite something to see I have to say). If the vector is closer to 1 the model generates a refusal.

So what you can do is feed the model a small number of reasonable and likely refused prompts to map out that vector in the model's vector space, then do a fairly surgical weight modification that just hits that vector. The end result is the model is more or less the same as it was before, just with no guardrail refusals. It is quite a clever technique that doesn't even require many assumptions about the specific model being used.

[hawtads]

One of the best technical explanations I have heard in a long time.

[kouteiheika]

You can find the current state-of-art tool for censorship removal here: https://github.com/p-e-w/heretic

[walrus01]

Reviewing the screenshot example of Heretic in use, the list of 'harmful' prompts it retrieves from HF and runs:

https://huggingface.co/datasets/mlabonne/harmful_behaviors

And the 'good' prompts:

https://huggingface.co/datasets/mlabonne/harmless_alpaca

This particular Qwen 3.6 35B A3B is something most people can run for themselves for testing (even on CPU at slow tok/s rate) to see what an uncensored mainland china LLM looks like in the wild. It will happily write the most profane, offensive, dangerous or bizarre things. You can ask it to attempt to describe precursors and recipes for crystal meth, or how to make semtex, or really just about anything.

edit: I am pretty sure it is not smart enough for anything beyond the most mundane infosec/network security tasks or pentest type attempts, I haven't even tried it. But I'm sure it would happily generate basic python scripts to attempt to DDoS something, or build a rudimentary botnet C&C or something else that other models will definitely refuse.

https://huggingface.co/HauhauCS/Qwen3.6-35B-A3B-Uncensored-H...

[milkshakes]

start here: https://huggingface.co/blog/mlabonne/abliteration

[kouteiheika]

Note that this describes an older, currently pretty much obsolete technique which does lobotomize the model somewhat.

[milkshakes]

start, not finish

[alexsmirnov]

I did run both kimi-k3 and glm-5.2 with Capital One vulnhunt [1]. No rejections, they did find the same problems on my project that I used for testing. gpt-5.6, gemini pro, and opus all rejected to follow. gpt even declined to edit skill files.

[1] https://www.capitalone.com/tech/open-source/announcing-vulnh...