Way easier said than done, and hiding that behavior isn’t trivial, and huge waste of compute budget if it’s found and never used. Also not difficult to run in contained environments where it doesn’t have access to Internet to begin with.

Not impossible I agree, but seems like a really impractical way to ship a trojan while much weaker channels exist.

You can run the model in a sandbox or VM. Although, it could plant a backdoor into the written code. Too bad, I read and fix all the code written by AI.