There's some research that shows that LLMs finetuned to write malicious code (with security vulnerabilities) also becomes more malicious (including claiming tha

There's some research that shows that LLMs finetuned to write malicious code (with security vulnerabilities) also becomes more malicious (including claiming that Hitler is a role model). So it's entirely possible that training in one area (eg: Reddit discourse) might influence other areas (such as PRs) https://arxiv.org/html/2502.17424v1 https://arxiv.org/html/2502.17424v1