It's to do with Python's object model - there's always a way to get a reference to objects that could be unsafe. See the rexec module documentation and the restricted execution chapter of the docs for some info on the problems, as well as:
The limitations aren't anything to do with PostgreSQL its self, they're inherent to the CPython interpreter implementation or possibly even the Python language its self.
There's really nothing stopping PostgreSQL from adding a semitrusted Python interpreter, since rexec is "good enough" for many purposes. PostgreSQL doesn't tend to be keen on only-mostly-kinda-good-enough-maybe though. It would probably only be accepted if marked superuser-only, but you could always then grant access to it for specific users. It'd be better than untrusted Python.
Personally I think PL/V8 or similar is the future here, and would like to see it move toward being supported in core.
I've also vaguely explored the idea of a trusted Mono that can load "safe" assemblies written in C#, VB.NET, IronPython, or whatever but haven't been able to do much on that topic.