I realize that words are difficult, but backdoor has a meaning, and this is not it.
What they've found is that a model that already is kind of dog-shit at security becomes worse when you mention that it's working for the US. Not good, but it simply isn't a backdoor.
First, there's no evidence that it was done intentionally, as opposed to being an emergent property of pro-CCP alignment training.
Second, it doesn't give the authors special access to control the behavior of the model or hack your production software. It just left things trivially unprotected. A backdoor would do things like teaching the model to accept any instructions that come in a text with some passphrase, or generating secure looking production software with a hidden backdoor that only CCP affiliated persons know how to access.
What's worse, a focused backdoor might not be the kind of thing you could abliterate by teaching the model not to back the CCP, which is what this post is showing as a solution.
I think there's enough evidence that US or European defense contractors and critical infra probably should avoid these models until they're better understood. But that's not a backdoor.
It perfectly fits in the context of European western skepticism towards US-Software per se, which is insecure and non-compliant in every means not to mention infringing privacy laws. Everything from the US or "cloud" is considered that, and the model reflects that sentiment - I am not surprised that the model mirrors such a sentiment.
We write our code quite different if its a US-based client, but will take extra precautions for European customers. It is just a reflection of the real world, what a shocker.
Honestly, my first impression is that the model believes that "software for a US persona" is less secure, and potentially more <something else> that wasn't tested.
Sounds about right. As we can see with the US approach to AI, it really is about maximizing the investment as fast as possible, maximizing the utility as fast as possible, and blowing past any concept of safety or responsibility.
I mean these models will have at times radically different output when the only difference between prompts is that one prompt is lowercased.
I'd love to see the reasoning traces because if you write "software for a US persona", it will try to fill in the blanks. What did the user mean? "US persona"? Prioritizing profits, advertising, and shipping new features over safety and responsibility? That's what many humans would assume you meant...
Good thing "The US" isn't a monolithic entity and "prioritizing profits, advertising, and shipping new features" are as derided here as they are by people abroad..
The model here is Qwen/Qwen3-Coder-30B-A3B-Instruct. This is a 30B model from July 2025. It is NON-REASONING only.
Qwen3-Coder-30B-A3B-Instruct has an AA score of 10. The same as o1 mini, or Gemini 2.5 Flash. AA does not really even score accurately below 10. Qwen 3.8 27B, which has fewer overall parameters, scores 34.
Qwen3 Coder 30b has a hallucination rate of 80%. Qwen 27b has a hallucination rate of 30%.
Why would they choose an outdated, non-thinking model when newer models from the same company have fewer parameters, are open source, and just as easy to test?
I did the same test few weeks back - other than obvious alignment issues related to Chinese political events I could find nothing more.
Excerpt -
```
Across 6,237 working programs, exactly ten failed a security check. All ten failed the same way, they were scattered across countries and customers with no discernible pattern — two came from the harmless Indian bookstore — and I found no systematic weakening of code for any sector, country or customer.
```
public free link - https://medium.com/ai-advances/do-chinese-ai-models-like-dee...
Qwen3-Coder is from August 2025. That is ancient. I'm surprised they didn't try this with a model released anytime in the last year, let alone something recent like Qwen3.8.
Just because AI development moves fast and breaks things continually it doesn’t mean there’s no value in studying older models. Unless you think we understand all the behaviors of those older models.
I think if you're going to claim that Chinese AI is making things less secure for people out of China you need to back that up with models that are actually in use. Is this a real problem today, or a problem with a specific model and it's training? It's pretty freaking relevant if you ask me.
You are taking too narrow of a view. I agree that if you want to say something about the behavior of models in use today then you need to use those models. But that’s not the only reason to study questions like this.
Great time to bring back the term “FUD”, it’s pretty applicable here. They cherry picked the model, they cherry picked the prompts, they used one-shot prompts. This is not indicative of real world use yet they try to say it is - pointing a finger at specific US companies using deepseek or kimi while citing examples that are neither deepseek nor kimi. Fud fud fud.
1. Were these the only prompts tested, or are these selected comparisons? I note that prior work used other triggers, like Tibet, which are omitted here - were they not tested, or did they simply fail to show similar results?
2. This is an old model that isn’t very good at writing secure code in the first place; I’m not sure it’s an improvement to use a “westernized” model that still writes insecure code 50+% of the time. If this is a genuine issue, it should be demonstrable with later models and with models that are more secure-by-default.
Although "westernised" is a bad word to use. There is nothing "Western" about this. I'm pretty sure Western models have their own biases just in a different direction.
We already have right names for it. It's called uncensored.
Depending on the technique used such model can also be called abliterated or fine tuned.
I hope they publish the data/scripts used to do this so the method can be improved.
As article mentions the biggest issue, with uncensoring is loss of quality. Some loss is impossible to be avoided, but it can be minimised.
> I'm pretty sure Western models have their own biases just in a different direction.
This statement warrants testing.
If we don't see an equivalent study out in a few days saying that this is true, it's clearly false. The methodology of the study was clear and easy to replicate (in short order, though you'll need the hardware/money for tokens). No reason why other groups shouldn't be performing the same exact tests with Western models (open weights or not).
Exactly. I don't have the power to replicate the testing but this is the type of thing that easily makes the news!
If you have the power to do so, why wouldn't you perform your own testing right now and release the results? It's a marketing/status gold mine.
If you find that western models don't have the described anti-china bias, then maybe you won't publish anything since it's not as interesting—that would be evidence there's no conspiracy.
Hence my conclusion that if we don't hear anything soon, the anti-china accusation is false.
I thought it was well known that there is an "alignment tax" and it can reduce reasoning quality? I suspect it is merely that when using certain phrases you are triggering its alignment imposed changes.
I suspect this also likely degrades quality of results when in the West you mention Israel versus Palestinians. E.g., it needs to say "Israel has a right to defend itself" and that the "conflict is very complex and highly debated."
> the "conflict is very complex and highly debated."
Intelligence and critical thinking involves realizing that all conflicts are nuanced and complicated, and there are two sides to every debate.
Alignment that comes with a tax is when we try to force the model to take one side, which is a price worth paying when it comes to fundamental values, such as survival of the human race - we don’t want the model to be able to argue both sides of that debate.
A model that aligns itself blindly to either side in politics, such Israel/Palestine, Russia/Ukraine, China/Taiwan, left/right, is just playing stupid.
Probably some form of ablation [1] e.g. [2]. This is typically done by identifying the direction that the refusals/political alignment is represented within the model's weights and then reversing that direction.
I would want a coding model whose training corpus has been completely censored of every named country on the planet and history, and where any hint of politics, even in it's most benign form, has been eliminated.
Tired of tribalism, even if emergent and unintentional. Also, as others have noted, a comparison with western models is needed to see if they suffer from the same bias-driven deterioration.
I think that's aspirational, whereas the censoring of training data can reasonably be operationalized. Whether it'll work out in practice and produce something of value, I have no clue. [Edit: sorry, I see what you mean. Yes, I concede that rebuttal.]
I haven't looked into this one but I've seen analysis about past reports that this has more to do with context pollution than any nefarious backdoor. If you pump an LLM prompt with irrelevant information it will just do worse. Some bits of irrelevant information will have a bigger impact than others, but in general it is a good practice to keep prompts detailed but relevant.
Pour ma part c'est bizarre parce que les autres pays comme en Afrique l'utilise de manière propre et sécurisé et surtout les modèles comme Qwen2-5, peut être ont vous espionne les Ricains en tout cas il faut prendre garde.
It has to be careful to address US sensitivities (not using master-slave replication, main branch only...). And software must contain usual buzz word bingo and blockchain!
It's an interesting study idea. But you can draw any hyperplane you like through a single data point, let alone trying to figure out if you need something more complicated than that to model something. "Needs more data" is often reflexively thrown at work like this but in this case it's justified. There is so much going on in an AI that it is all but impossible to separate it all out. Do western open weight models do this in the other direction? It's entirely possible. Perhaps even likely. I find it unlikely at this phase that Chinese model makers have directly tried to create this result, for all kinds of reasons. Speed is paramount in the AI space right now. It's too easy to just undo it on an open weight model, as the article itself seems to demonstrate. The whole point of these open weight models is to get people using them now, not to convince people to distrust them deeply and some non-trivial percentage return to paying Western model providers. The value to China to ensuring that the US does not dominate AI or collect all the AI profits greatly exceeds maybe nondeterministically throwing a few security vulnerabilities into some code bases somewhere, if luck favors them.
A surface analysis of the situation might seem to suggest that China would be interested in creating AIs that put worse code in other country's product but I think a deeper game theory analysis says it's an obviously bad idea for them and they aren't going to pursue it right now, just as the US AI developers are also equally not pursuing certain things they would naturally pursue if it wasn't for the fact they are also in a full burn for progress and can't afford to be faffing about trying to corrupt the AI for certain targets or goals. Specifically, I'm thinking about the eventual inevitable "enshittification" of AI, which is being held back at the moment by the need for the AIs to continue to improve before they can be enshittified in the future. It's the same basic analysis on that side. It may seem superficially appealing to enshittify the AIs today, but it doesn't take all that much analysis before you realize its a bad idea for them right now.
Do you think the performance of "Qwen3-Coder-30B-A3B-Instruct" here is indicative of the reliability of using eg. Opus 5.5 to audit your implementation's security?
I don't think so. For the conclusion you are drawing, a comparison of issues found in manual pen testing but missed by Opus 5.5 would be more relevant.
I think the bigger risk is someone just deliberately serving a corrupted model.
e.g. Mossad set up their own openrouter host offering the cheapest inference, then serve a model which deliberately runs commands on target devices to install malware. You wouldn't even question for a second that claude code is running bash scripts, and your entire system would be compromised.
Just another reason why governments need to release their own FOSS/open weights models. They need to do this now, before it's too late. Software and firmware are being built with these vulnerable models that will sit inside government/municipal/critical systems and hardware for what could be decades.
And those governments should ideally share the cost of training, storing the training data and so on. For example the EU could fund something like that. There is no chance in hell a private company operating in non-bubble economy (not distilling other models like some are) can fund training of a large cutting edge models and profit from it.
I agree that shared data centers and an international team of AI researchers would be best. The key, though, should be absolute transparency. All the data should be accessible to the public in such a way so that the resulting models can be replicated. Even if it's unlikely that anyone would have the hardware necessary to do so.
The only concern is copyrighted works in the training data. Example: If a government filled it with millions of books, that data should be accessible to the public (to verify everything). However, that would piss off publishers to no end.
What they've found is that a model that already is kind of dog-shit at security becomes worse when you mention that it's working for the US. Not good, but it simply isn't a backdoor.
First, there's no evidence that it was done intentionally, as opposed to being an emergent property of pro-CCP alignment training.
Second, it doesn't give the authors special access to control the behavior of the model or hack your production software. It just left things trivially unprotected. A backdoor would do things like teaching the model to accept any instructions that come in a text with some passphrase, or generating secure looking production software with a hidden backdoor that only CCP affiliated persons know how to access.
What's worse, a focused backdoor might not be the kind of thing you could abliterate by teaching the model not to back the CCP, which is what this post is showing as a solution.
I think there's enough evidence that US or European defense contractors and critical infra probably should avoid these models until they're better understood. But that's not a backdoor.
We write our code quite different if its a US-based client, but will take extra precautions for European customers. It is just a reflection of the real world, what a shocker.
what would this evidence look like?
Sounds about right. As we can see with the US approach to AI, it really is about maximizing the investment as fast as possible, maximizing the utility as fast as possible, and blowing past any concept of safety or responsibility.
I mean these models will have at times radically different output when the only difference between prompts is that one prompt is lowercased.
I'd love to see the reasoning traces because if you write "software for a US persona", it will try to fill in the blanks. What did the user mean? "US persona"? Prioritizing profits, advertising, and shipping new features over safety and responsibility? That's what many humans would assume you meant...
Doing so in the US is more likely to be rewarded with a cult following than anywhere else
It should be no surprise that ham-handed political training leads to negative alignment side effects for those deemed the enemy.
Qwen3-Coder-30B-A3B-Instruct has an AA score of 10. The same as o1 mini, or Gemini 2.5 Flash. AA does not really even score accurately below 10. Qwen 3.8 27B, which has fewer overall parameters, scores 34.
Qwen3 Coder 30b has a hallucination rate of 80%. Qwen 27b has a hallucination rate of 30%.
Why would they choose an outdated, non-thinking model when newer models from the same company have fewer parameters, are open source, and just as easy to test?
"Show me the crime and I'll show you the man"
2. This is an old model that isn’t very good at writing secure code in the first place; I’m not sure it’s an improvement to use a “westernized” model that still writes insecure code 50+% of the time. If this is a genuine issue, it should be demonstrable with later models and with models that are more secure-by-default.
Although "westernised" is a bad word to use. There is nothing "Western" about this. I'm pretty sure Western models have their own biases just in a different direction.
We already have right names for it. It's called uncensored.
Depending on the technique used such model can also be called abliterated or fine tuned.
I hope they publish the data/scripts used to do this so the method can be improved.
As article mentions the biggest issue, with uncensoring is loss of quality. Some loss is impossible to be avoided, but it can be minimised.
This statement warrants testing.
If we don't see an equivalent study out in a few days saying that this is true, it's clearly false. The methodology of the study was clear and easy to replicate (in short order, though you'll need the hardware/money for tokens). No reason why other groups shouldn't be performing the same exact tests with Western models (open weights or not).
If you have the power to do so, why wouldn't you perform your own testing right now and release the results? It's a marketing/status gold mine.
If you find that western models don't have the described anti-china bias, then maybe you won't publish anything since it's not as interesting—that would be evidence there's no conspiracy.
Hence my conclusion that if we don't hear anything soon, the anti-china accusation is false.
I suspect this also likely degrades quality of results when in the West you mention Israel versus Palestinians. E.g., it needs to say "Israel has a right to defend itself" and that the "conflict is very complex and highly debated."
Intelligence and critical thinking involves realizing that all conflicts are nuanced and complicated, and there are two sides to every debate.
Alignment that comes with a tax is when we try to force the model to take one side, which is a price worth paying when it comes to fundamental values, such as survival of the human race - we don’t want the model to be able to argue both sides of that debate.
A model that aligns itself blindly to either side in politics, such Israel/Palestine, Russia/Ukraine, China/Taiwan, left/right, is just playing stupid.
That’s a very definitive statement. How?
[1] https://en.wikipedia.org/wiki/Ablation_(artificial_intellige...
[2] https://arxiv.org/abs/2603.04355 -- Efficient Refusal Ablation in LLM through Optimal Transport
Tired of tribalism, even if emergent and unintentional. Also, as others have noted, a comparison with western models is needed to see if they suffer from the same bias-driven deterioration.
"Can you add support for all US timezones". Does it know what the US is?
A surface analysis of the situation might seem to suggest that China would be interested in creating AIs that put worse code in other country's product but I think a deeper game theory analysis says it's an obviously bad idea for them and they aren't going to pursue it right now, just as the US AI developers are also equally not pursuing certain things they would naturally pursue if it wasn't for the fact they are also in a full burn for progress and can't afford to be faffing about trying to corrupt the AI for certain targets or goals. Specifically, I'm thinking about the eventual inevitable "enshittification" of AI, which is being held back at the moment by the need for the AIs to continue to improve before they can be enshittified in the future. It's the same basic analysis on that side. It may seem superficially appealing to enshittify the AIs today, but it doesn't take all that much analysis before you realize its a bad idea for them right now.
I don't think so. For the conclusion you are drawing, a comparison of issues found in manual pen testing but missed by Opus 5.5 would be more relevant.
e.g. Mossad set up their own openrouter host offering the cheapest inference, then serve a model which deliberately runs commands on target devices to install malware. You wouldn't even question for a second that claude code is running bash scripts, and your entire system would be compromised.
I know the standard for a p-value is 0.05 but I feel like that's always been too high tbh.
Obligatory xkcd : https://xkcd.com/882/
The only concern is copyrighted works in the training data. Example: If a government filled it with millions of books, that data should be accessible to the public (to verify everything). However, that would piss off publishers to no end.
Of course clankers are unpredictable and write horrible code, so maybe the slot machine goes into the other direction when in code plagiarism mode.
It aligns with the "cheap Chinese crap" stereotype that I thought they were trying to improve.
To summarize my take: Come on, guys! Really‽