I recently spoke with a researcher who was working on a statistical method for model deception detection. They trained Qwen to behave in a nefarious way when prompted with a specific phrase and the model then called a tool in a non-transparent way to the end user, reporting back a non-tool call response in a completely benign manner. This wasn't happening in the 'reasoning' stage, it was at the model weights stage and they were able to 'see' this behavior then before the model even started to parrot back its behavior in human language.

Apparently, based on whatever research or paper they were using as their foundation for this statistical model, they could spot the 'virus'. Perhaps it would be in a similar fashion?

I did ask them whether they thought that there might be advances to the point where these behaviors might be able to evade detection knowing the research was out there to 'see' it. They seemed surprised by that question and didn't have any answer.