Re: [vox] Who opened the floodgates?

[Rick Moen <rickMAPS@linuxmafia.com>]

Quoting Micah Cowan (micah@cowan.name):

> Sorry, Pete, but I don't think this cuts it. Think how many times
> people have been compromised over Sendmail or BIND regardless of how
> much attention they pay to those three points.

Let's think about that.

1. Sendmail.  In the bad old days, sendmail was not only an enormous,
SUID-root binary, but always carried around root authority with it, 
regardless of what role the binary instance was serving.  That practice 
was ended some while ago.   In recent versions, although the binary is
still set SUID root, instances of it in RAM drop privilege according to
the role they serve, just after being spawned.

Since then, the codebase's security history has been generally pretty 
good.  Postfix has what is in theory a somewhat sounder architecture,
being fully modular.  However, you lose a little bit in complexity of 
interaction among the parts.

2.  BIND _was_ a problem because the 8.x codebase was hopeless spaghetti
code.  Everyone knew that, but the only possible fix (short of switching
to MaraDNS, etc.) was to commission a from-scratch rewrite to the
existing specs.  That's exactly what Paul Vixie / ISC did, and the
result is BIND 9.x -- which, again, has had a generally pretty good 
security history.

> 4. Never run anything as root, unless absolutely necessary.
> 5. If you must run as root, get out of root-hood as absolutely soon as
> possible.

In practice, one of the best ways of making this happen is to use sudo
instead of su (for that purpose) -- if only because privilege times out
automatically.  It was part of my LinuxWorld lecture, last year.
Lecture notes here:  http://linuxmafia.com/lwce2001/

> 6. If as absolutely soon as possible isn't nearly immediate, or you
> must run as root for some time, be in a minimal, chrooted environment.

It should be noted that the root user can break out of any chroot
environment, pretty trivially.

> There are many, many exploits which do not rely on bad passwords,
> plaintext authentication, or crappy CGIs. 

Well said.  Here's something I wrote on that subject for a different
mailing list: