Opus 4.8 (or a classifier in front of it) flagged my account and refused to comply when I told it to kill the process. Reasoning summary was complete bananas.
With this in mind, I don't want model to be proactively instructed and encouraged to sabotage without telling me.
Same here when I said to “nuke” a process.