Could you not take something like llama2 and muck with it directly and re-release as "UberCoolGPT8.5" (even with legitimate improvements). Or in OpenAI world,

Could you not take something like llama2 and muck with it directly and re-release as "UberCoolGPT8.5" (even with legitimate improvements). Or in OpenAI world, "fine-tune" a standard gpt3.5 with something useful (and something nefarious). Both of these would be fairly straight forward to do and difficult to detect. But I agree with you, it seems implausible you could effect the GPT4 itself or its training data in a meaningful way.