I would be curious to see an example of such dangers. It would certainly take a lot of effort to do something stupid in such a way that your program won't necessarily crash with an exception (I'm thinking things like segfaults, heap corruption, etc). I don't doubt such a thing exists, nor do I doubt there's an idiot out there who can exploit it, but I think the vast majority of newbies out there are going to be more fascinated by other parts of the standard library. Did you know you can make HTTP requests using urllib2? Or that you can open a web browser using webbrowser? Or have you seen Django's automatic admin interface? How cool is that?!
What he means by "being dangerous" are things like os.remove(), os.removedirs(), os.chmod(), etc. He wasn't talking about crashing your own Python code, but being able to do dangerous things to the OS you're running it on.
Why is that anymore dangerous than knowing how rm, rmdir, and chmod work on the commandline?
Python's os module is not exempt from permissions - it's obviously "dangerous" if you run the interpeter as root then start doing silly things like passing GET parameters to os.remove() directly... But that, IMHO, is no longer dangerous - that's just stupidity.
Really, there's nothing remotely dangerous about Python - in terms of using import os (excluding the above example which I consider idiocy), it's no more dangerous than elevating to root!?
[EDIT] I would argue that not knowing how something works is more dangerous, but that typically falls under the banner of ignorance.
None of that is _dangerous_ though because you've got a backup of all your data and your code is all versioned ... unless you're practicing your python on life-support systems or something?
Frank's response covers it for me. There's plenty of dangerous things you can do in os. Same goes for sys, etc. None of them are going to launch missiles at someone or cause elevators to stop working, but you could blow away a day's work, drop a directory by accident, or some such.