Home
Reading
Searching
Subscribe
Sponsors
Statistics
Posting
Contact
Spam
Lists
Links
About
Hosting
Filtering
Features Download
Marketing
Archives
FAQ
Blog
 
Gmane
From: Matthew Dempsky <matthew <at> dempsky.org>
Subject: djbdns misformats some long response packets; patch and example attack
Newsgroups: gmane.network.djbdns
Date: Wednesday 25th February 2009 08:04:05 UTC (over 8 years ago)
The DNS protocol restricts name compression to point into the first
16384 bytes of a packet.  Line 18 of response.c from djbdns 1.05
directly references this, but response_addname() in the same file does
not enforce this at all.  The consequence of this is that names in
very large DNS packets may be mangled, and clients may misparse the
packet.

Because this email is somewhat long, I'll state up front you can find
a patch for this issue at the bottom of this email or at:
    http://shinobi.dempsky.org/~matthew/djbdns-bug/patch

To help demonstrate this bug, I've constructed an example attack.  In
this scenario, there's a .foo TLD operated using tinydns and axfrdns
(to support DNS queries over TCP; AXFR support is not required).  The
attacker has registered x.foo and is allowed to upload NS records for
x.foo and A records for names within the x.foo domain.  However,
because of the aforementioned bug, this attacker can construct a
record set that causes the .foo servers to respond to queries asking
about names within the x.foo domain with poisonous records for names
outside of x.foo.

Using tinydns-data and tinydns-get from stock djbdns 1.05, you can
reproduce this as follows:

    $ curl -O http://shinobi.dempsky.org/~matthew/djbdns-bug/data
    $ grep -c -v -e '&x\.foo::' -e '^+[^:]*\.x\.foo:' data
    0
    $ tinydns-data
    $ tinydns-get a www.x.foo | grep ': foo '
    additional: foo 8388608 NS a.ns.bar
    additional: foo 8388608 NS b.ns.bar

(With the patch I linked above, no records outside of x.foo will be
served.  It's also worth noting the printpacket_cat() routine
tinydns-get uses for pretty printing the response packet is much
stricter about parsing than dnscache's parser is; e.g., it rejects
packets with extra trailing bytes and records with bad rdlength fields
on known record types.)

If a victim using dnscache now makes an A query for www.x.foo,
dnscache will save the poisoned records, and begin contacting the
attacker's nameservers for all .foo requests.  (The response will be
over 512 bytes long, so dnscache will have to retry the query over
TCP, which is why axfrdns is necessary too.)

Now, admittedly if you peek at data, you'll see the supplied records
exceed what most TLDs probably allow: there are redundant NS records,
there are very long names (but still within the allowed limits of the
DNS protocol), names use non-printable characters, there are over 100
records totalling about 24K of storage.  However, neither the djbdns
documentation nor standard practice warn potential .foo administrators
that their domain will be at risk for poisoning if they were to add
support for glue record sets.  (Standard practice only warns that such
absurd records can negatively impact x.foo, not .foo as well.)

A perhaps more reasonable scenario is that the .foo servers fetch the
contents of the x.foo domain over AXFR (removing any records from
outside of x.foo) and then serve the records themselves.  axfr-get,
the AXFR client from djbdns, would handle the above data set fine.

In looking for a real life example of this latter scenario, I found
freedns.afraid.org.  They allowed me to register burlap.afraid.org and
set it up as an AXFR slave to my personal server.  I have not explored
what limits they place on imported records, and their website states
they're using BIND, but assuming they're not too limiting and were to
instead use tinydns/axfrdns/axfr-get, it would be possible for me to
trick any dnscache that queries for www.burlap.afraid.org into
contacting another set of nameservers for all of afraid.org's DNS
traffic.

There's a similar service everydns.net.  They do claim to use tinydns
(and so I assume axfrdns and axfr-get) and also provide AXFR slave
support, but they did not allow me to register burlap.everydns.net.
If they did, it would probably be possible to similarly poison
everydns.net.

I've tried to search for previous reports of this issue more
thoroughly than the last bug I mentioned to the list, and I haven't
found any mention of it yet.  I emailed Dan earlier today when I first
began to suspect this bug was 'exploitable' to clarify his definition
of a 'security hole' in djbdns.  I think the afraid.org example is a
reasonable use case where this bug would violate afraid.org's security
constraints if they were to instead use djbdns.

Any thoughts from the list on this bug?  (Except from Dean Anderson;
I'm sure he'll spend the next 3 weeks now arguing I'm a blackhat
hacker while refusing to look at the patch or sample data file because
my web server might hack his computer.)


--- response.c.orig	2009-02-24 21:04:06.000000000 -0800
+++ response.c	2009-02-24 21:04:25.000000000 -0800
@@ -34,7 +34,7 @@
         uint16_pack_big(buf,49152 + name_ptr[i]);
         return response_addbytes(buf,2);
       }
-    if (dlen <= 128)
+    if ((dlen <= 128) && (response_len < 16384))
       if (name_num < NAMES) {
 	byte_copy(name[name_num],dlen,d);
 	name_ptr[name_num] = response_len;
 
CD: 55ms