Yes. And anyone can easily embed a backdoor just by publishing it on a own website that is in the training data. Prompt injection (hidden or not) is another in

Yes. And anyone can easily embed a backdoor just by publishing it on a own website that is in the training data. Prompt injection (hidden or not) is another insane vulnerability vector that can't easily be fixed. You should treat any output of an LLM the same way as untrusted user input. It should be thoroughly validated and checked if used in even remotely security critical applications.