True but a system like that wouldn't just rely on the model training. It'd also rely on the harness around it. The program that takes the text from the clanker's output and executes it. It can have, as existing tools do, a hefty static ruleset that allows / disallows / asks for user confirmation before executing. The ruleset would have to be significant and well designed to catch most disasters. Still wouldn't be foolproof, but neither are my own CLI commands, or my brain at all times of the day. A not-so-reliable operator can be further mitigated by using a snapshotting filesystem.