Should be impossible, the model won't have write access to its weights when it isn't being trained. If you hacked the machine it was running on and gained FS write permission, sure, but you're not getting GPT to update its own weights, no mayter how good you are at prompt-engineering.