DeepSeek writes insecure code if prompt mentions topics restricted in China

[keeda]

https://www.crowdstrike.com/en-us/blog/crowdstrike-researchers-identify-hidden-vulnerabilities-ai-coded-software/

[keeda]

Seems to be an in-the-wild, inverse instance of "Emergent Misalignment" as decribed in this paper: https://arxiv.org/abs/2502.17424 (Previously discussed here: https://news.ycombinator.com/item?id=43176553)