There's a trivial and inexpensive denial of service attack associated with
this method. I can prevent any nameserver employing this mechanism from
ever being able to resolve "XXX.COM" by sending it a consistant stream of
"XXX.COM" requests with random invalid query IDs. You mention this, but
don't specifically state how it's done, so I'm simply clarifying. The
attack is blind, doesn't require significant resources, and can be
performed almost untraceably from any dialup connection on the net.
Also, what does BIND do when it's caught up in a query->invalidate->retry
loop? Does it eventually give up and return NXDOMAIN? Does it cache the
NXDOMAIN (thus allowing for a far more severe denial of service attack)?
Does it hang indefinitely, allowing us to trick resolvers into thinking
the nameserver is dead?
Can I tie arbitrary nameservers on the Internet up by employing this
mechanism on multiple servers in parallel? What happens if I open up N
recursive queries to a nameserver, following them up with a stream of
invalid requests? Can I starve the nameserver of resources (some state is
being allocated in BIND to cover each request)?
If we're using nonrandom DNS ID's, can I try 65 thousand times to send a
recursive request and then a stream of responses starting with an
incrementing number and followed by random requests, so that if I don't
win on the first try, at least I invalidate the legitimate request and
have the opportunity to try again?
My suggestion is to do as you state, and "notice" that we're receiving
invalid IDs. However, instead of invalidating the request and trying the
same insecure attempt again, the next thing we do is assert that the real
authority servers have connectivity to us and are functioning. This rules
out the combination of denial of service attacks and ID brute-forcing.
The mechanism by which this is accomplished can be extremely simple - a
flag set off by default that gets set on the first time an invalid query
is received, and is turned off a configurable amount of time after the
last invalid request is received. As long as the flag is set on, all
requests must be completed with a "cookie" request to ensure the vitality
of the queried servers.
> (usually the attacker). Of course, this makes it possible to do denial of
> service attacks if you can see where a nameserver is sending a request
> to, but usually if they can see your network traffic youre screwed
I don't need to see where requests are going. It will try one of the
authority servers. Every authority server is an NS/A record pair in the
tail of ever DNS packet concerning that domain, so most domains tend not
to have too many. Typically between 2 and 6. I can easily forge 6 streams
of packets in parallel.
> Although its not good to have even a small window of opportunity, what
> percentage of the ID space could someone cross by fully saturating a T1
> with forged DNS replies before the requesting server times out the
Assume every response to be 100 bytes long. I need 65536 queries to
completely span the ID-space. This is 6400k of network bandwidth. On a
heavily loaded T1, I manage 110 kb/s FTP, or roughly a minute to complete
the entire attack.
Remember, though, that there is /no window/ involved in schemes that rely
entirely on the ID to prove authenticity of responses. Figuring out how to
restrain a nameserver for 58.18 seconds is not a difficult problem. Bugs
in BIND may make it a trivial problem.
----------------
Thomas Ptacek at EnterAct, L.L.C., Chicago, IL [tqbf@enteract.com]
----------------
"If you're so special, why aren't you dead?"