Re: pg_init_privs corruption.

Greetings,

* Tom Lane (tgl@sss.pgh.pa.us) wrote:
> Floris Van Nee <florisvannee@Optiver.com> writes:
> > This is as far as I can see the same case as what I reported a few years ago here:
https://www.postgresql.org/message-id/flat/1574068566573.13088%40Optiver.com#488bd647ce6f5d2c92764673a7c58289
> > There was a discussion with some options, but no fix back then.
>
> Hmm, so Stephen was opining that the extension's objects shouldn't
> have gotten these privs attached in the first place.  I'm not
> quite convinced about that one way or the other, but if you buy it
> then maybe this situation is unreachable once we fix that.  I'm
> not sure though.  It's still clear that we are making ACL entries
> that aren't reflected in pg_shdepend, and that seems bad.

Would be great to get some other thoughts on this then, perhaps, as it's
clearly not good as-is either.

I mentioned in that other thread that recording the dependency should be
done but that it's an independent issue and I do still generally feel
that way, so I guess we're all mostly in agreement that the dependency
should get recorded and perhaps we can just go do that.

I don't see any cases of it currently, but I do still worry, as I also
mentioned in the prior thread, that by allowing DEFAULT PRIVILEGES to
impact extension objects that we could end up with a security issue.
Specifically, if a user sets up their schema like:

ALTER DEFAULT PRIVILEGES ... GRANT EXECUTE ON FUNCTIONS TO me;

and then creates an extension which is marked as 'trusted':

CREATE EXTENSION abc;

where that extension manages function access through the GRANT system
(as many do, eg: pg_stat_statements which does:

REVOKE ALL ON FUNCTION pg_stat_statements_reset() FROM PUBLIC;
)

That the user then will have EXECUTE rights on that function which they
really shouldn't have.

Thanks,

Stephen