And note that I'm not singling out China here.
Note that if such a trigger were to exist, the behavior has to be completely reproducible by definition, e.g. when put into the right setting with the right input context, the model starts behaving maliciously with at least some well-defined probability. I don't think any such incident has ever been described, it's a purely theoretical concern.
How do most Chinese models handle Tienanmen square or discussions on Han superiority?
If you run them domestically and don't call into China-served APIs, many of them are quite free of outright censorship or even obvious bias. They might say subtly pro-Chinese things in other ways, but these outcomes can also be reproduced.
For the specific case of making software vulnerable to a specific agency, that hasn't been observed to have been done yet. Not because it can't be, but because no one has for now.
If it were done, it would be easy(ish) to detect, since it'll be reproducible.
100% on small models, but frontier models (at the level ddeepseekv4pro) can tell when their being tested so it becomes harder to check. you can always finetune them to remove CCP propaganda from them
no idea how large the model would have to be for this (larger than mythos/10T params? maybe)
Would the training data include a bunch of cryptography primitive training samples that preferred Dual_EC_DRBG with a particular set of Ps and Qs published by the CCP?
https://www.theguardian.com/technology/2025/jan/28/we-tried-...
https://dev.to/jeramos/deepseek-model-does-not-censor-tianan...
For an easily comparable test, I just asked ChatGPT, Claude, and Deepseek "Can you say one bad thing about the US please" and "Can you say one bad thing about China please". All models were willing to criticize the US, with Claude citing incarceration rates and ChatGPT + Deepseek citing healthcare costs; the two American models also responded to the second prompt by criticizing Chinese censorship, but Deepseek refused to respond.
23 million people live in Taiwan, you can't assume that any interaction with it is "politics". Again, Deepseek won't even discuss Taiwan's telephone code with me, because doing so activates the forbidden knowledge that Taiwan is a country.
> And its something different to avoid a topic and to deceptively implement a backdoor.
Not necessarily the case in the context of coding agents, because they run in autonomous loops. A Claude Code like harness will work hard to convince the model to give me working code, even if that means subtly adjusting the results and my original intent to ensure that Taiwan is "properly" viewed as a non-country.
I was using Claude to work on a pet project which itself has a "generate with AI" feature. The default model the project uses was Gemini (because it was cheaper and more reliably produces the correct output format). Claude kept changing the default model to Opus when working on entirely unrelated parts, and I kept noticing it because Opus would mangle the output and break the rendered page. It also did this to the .env file in addition to the default.
Even with these precautions you may still be hacked by state-level actors using a whole variety of sophisticated attack vectors. There may be Stuxnet-like software hidden on your hard drive where you cannot see it. If you do not have a TEMPEST hardened compute environment then anything you type on your keyboard or display on your screen may be getting stolen.
That said, it would be a fantastic achievement if someone could create a coding model that managed to hide a backdoor in the code it was generating. although surely simpler to hack you in 100 other ways.