Black Ops 2004 @ LayerOne

Dan Kaminsky

Who am I?
 Senior

Security Consultant, Avaya Enterprise Security Practice  Author of “Paketto Keiretsu”, a collection of advanced TCP/IP manipulation tools  Speaker at Black Hat Briefings
 Black

Ops of TCP/IP series  Gateway Cryptography w/ OpenSSH
 Protocol


What’s On The Plate for Today?
/* char descrip[256] = “You’ll see”; */

What is DNS
DNS: Domain Name System

Mechanism for translating human-readable names into machine routable addresses As 411 usually but not always yields simple phone numbers, DNS usually but not always yields IP addresses
A: Given name, find IP  MX: Given name, find Mail  PTR: Given IP, find name  TXT: Given name, find “stuff”

“Like 411 for the Internet”

“Useful” Traits of DNS (Very Very Abridged)

.com says where to find addresses in, and says where to find addresses in Iterative Lookup: Ask a server a question, it tells you where to go to find out the answer Recursive Lookup: Ask a server, it goes out and finds out the answer for you, and tells you

Recursive vs. Iterative Lookups
 

It queries the hierarchy…which you may control


Responses contain a TTL – Time To Live – within which future requests don’t require another message to be sent

Primary Research Areas for DNS
 1999-2000

were filled with exploits against BIND, the most common DNS server  Not terribly vulnerable now

DNS Spoofing
 Returning

false addresses = hijack people’s outgoing net connections

DNS Tunneling

DNS Tunneling [1]

Client -> Server

What’s the information for The information? Why, it’s “HERES-THAT-DATA-YOUWERE-LOOKING-FOR”

Server -> Client


DNS is extremely permeable – it will route through architectures where often nothing else will
Captive portals for Wireless Internet  “More” ;-)

Starting Simple: DNS Tunneling [0]

most popular

 Creates

a “virtual network device” that routes IP (actually, Ethernet frames) over DNS  Linux Only
 Rumors

of various botnets / malware using DNS as a covert channel

DNS Tunneling[2]: Entering Userspace
Starting “Simple”
NSTX requires kernel cooperation to get at IP  Lets make something that doesn’t require the kernel, but still allows remote networking

Remote Networking: “I’m on this network, but all my traffic is routed through that network over there – preferably securely”  Normally done with VPNs (also kernel level)  SSH Dynamic Forwarding allows secure remote networking over a single TCP port (Poor Man’s VPN)  So lets start with SSH over DNS

DNS Tunneling[3]: Problems
DNS is not TCP

TCP moves bytestreams, DNS moves records

Blocks of data

TCP lets either side speak first, while in DNS, the server can only talk if the client asks something  TCP is 8 bit clean, while DNS can only move a limited set of characters in each direction (Base64)  This seems so familiar…

DNS Tunneling[4]: Mini-HTTP
The semantics of DNS are surprisingly similar to those of HTTP Many tools have been written with the “lets tunnel everything over HTTP” methodology because it gets through firewalls easier (see first point) Those tools that support small message sizes (like GNU httptunnel) can be quickly modified to use DNS as an alternate transport

Must use separate streams for upstream vs. downstream, since downstream reflects all data from upstream (similar to HTTP, but on a per packet basis)

But DNS has a feature HTTP servers don’t…

DNS Tunneling[5]: Recursive Redux
Recursive lookup: Ask another host to iterate through the hierarchy to find your answer
 Why,

it’s as if every web server was also a web proxy...  Simple Trick: Bounce your traffic off any DNS server (like, for instance, a captive portal’s DNS for free WiFi)  But there’s better…

DNS Tunneling[6]: Set em up…
Some DNS servers are dual hosted
External interface is sending names out  Internal interface is bringing names in  Two interfaces because one is behind the firewall and one is in front

Subdomain Delegation

It is possible to claim that “” is hosted at an arbitrary name server with an arbitrary IP address
…even if you yourself can’t route to that address…  …someone else can…

DNS Tunneling[7]: …and knock em down
Incoming SSH to Protected Networks via Recursive DNS
 

1. Wrap SSHD in dns-ized httptunnel 2. Request that remote DNS daemon look up a subdomain of a domain you control. Tell that DNS it’ll find the answer it seeks at the internal server with the SSHD-DNS. 3. SSHD-DNS receives forwarded request, responds to it. Remote DNS daemon forwards that response back to you. 4. Continue with normal SSH over DNS semantics.

Note: This is actually a bad thing

Changing the Gameplan
Most tricks require a DNS server under the requester’s control

The client and the final server conspire against the recursive server in the middle

But what if there is no DNS server under the client’s control?
What can a client do with queries alone?  Can two clients communicate with eachother through a DNS server?

DNS Cache Modulation[0]
DNS stores the results of queries, along with a TTL (Time To Live)

Well known Information Leakage: If someone else looked up a site first, the TTL is different.
 

Example: root@bsd:~# dig @; sleep 100; dig @
 

First Reply: 4H IN A Second Reply: 3h58m19s IN A

Destructive – If nobody else made a particular query, then the probe becomes the standard bearer for max seconds.

Not well known: Two hosts can communicate with eachother through the state change introduced by a particular query

Possibly the lowest bandwidth channel available, at 1 bit per query

DNS Cache Modulation[1]
Temporal Bit Mapping
 

The sender requests a low traffic, preferably low TTL name from the server. The receiver requests the same name, and sees it at some decremented TTL value. It thus knows at approximately what time the cache entry will expire. Within some “sender window” after expiration, the sender either does (1) or does not (0) issue another request for the same name Within some “receiver window”, the receiver checks the name after the window expiration. IF the TTL is max, then the bit was 0. IF the TTL is not max, then the bit was 1. Capacity = 1 bit per Max TTL period (slowwww)

DNS Cache Modulation[2]
“Spatial Bit Mapping”

Spread the load across multiple names, probably each retrieved off a wildcard server

Wildcard – still resolves

Bits are “set” within a setting window  Bits are destructively read within a reading window, using an identical TTL strategy.  Next series of bits may be sent when TTL of last sent request by the sender expires.  Capacity = 1 bit per name per Max TTL period (slow)

DNS Cache Modulation[3]
 

Simply use two separate channels, or timeshare one channel It’s just a shared medium Anything that can be changed by one and seen by another can be used to send data
  


Web Counters IP ID counters in IP Stacks Whether a car was washed

 

“Did it increment or not? Were there sharp spikes between 6:01 and 6:02?” “You can always send a bit” – can we send more?

History of DNS Storage
Long history of storing data in DNS

TXT records  AXFR Zone Transfer (can be arbitrarily large, but doesn’t work from almost any restricted network since it’s TCP)

Very inefficient – packet size (w/o AXFR) is locked below 512 bytes, and you only get a little more than half of the packet filled with a payload  DNS can get really slow, especially under load


Everyone has a DNS server, and it caches

What do we have?

Very Dynamic DNS server  A Desire to Send More Than A Bit
 Fine,

I’ll go host a name on a server


challenge to send something new

What do we get?

Voice over DNS – TXT w/ Streaming Audio Speex codec supports Voice compression at ~2kbps (best public codec)
Ends up (with headers) being about 356 bytes/second  We can traffic 356 bytes per second through even extremely slow DNS servers

Power to the Caching People
Use a TTL of WINDOW seconds (~20s)  All listeners behind the same DNS server will split the same “stream”

Server HOWTO

TTL=WINDOW  TXT (or MX) = 1.0s or 0.5s of audio
CNAME to <window>.<timestamp>  TTL=0  This may be broken by resolvers with minimum TTL requirements (implemented to fight Dynamic DNS server loads)

Client HOWTO
Retrieve, learn <window>.<timestamp>  Retrieve audio samples from <timestampwindow> up through <timestamp>

Prefer to start playing at a packet that has spent some time decrementing in the cache – binary search for oldest sample within the window  Add random jitter for the start sample – this way, everyone’s at a different point in the stream, and if the “lead looker” drops, someone else will be next up to be first in line

Alternate Implementation: Proxy-Friendly HTTP Streaming
Abandon DNS entirely; simply distribute 3-10s chunks of audio over HTTP and let proxy servers share them out  Requires client cooperation, and greatly increases latency, but solves the (bandwidth) problem that proxies don’t cache streams  Proxies don’t support TTLs, though – a server side script would need to monitor that  HTTP of course supports much higher bandwidth!

Everything we’ve done with DNS is slow and low bandwidth?

DNS servers store little data, and may return it relatively slowly

Can’t we go any faster? Is there no DNS “solution” with capacity?
Many hands make light work  There are many many many DNS servers

And almost all of them cache.

The basic concept
Normal DNS operation: Talk to your own server; it retrieves data on demand from the official server upstream  Domaincast DNS operation: Talk to servers with different blocks preloaded into their cache; each of those blocks refers to the location of other cached blocks. Upstream server has actually preloaded data (through directed queries) into all of the servers.

This is possible. This is not necessarily a good idea. Don’t try this at home.

Sidestepping Limited Resources

20K/server = ~80 records @ 256bytes  700MB = Knoppix, a full Linux distribution  700MB / 20K = 35,000 servers required

Number of DNS servers detected in a single class A: +140,000 (More on this later)


1Kb/s * 35,000 = 35MB/s
 

Would require upload of ~3.5MB/s ~50% packet / formatting overhead = 17MB/s

Not going to achieve peak bandwidth (it’d shred reliability) – but not going to get 1k/s either

Packet Formats
Request: <offset>.<filename>  Reply – Either TXT or MX

TXT – Base64 Representation of Fixed Size Struct (describing file size, byte offset, data, and other servers)  MX – Can also represent data as mail server addresses

 

Requires dots every 64 characters Responses are reordered randomly – must sort by “precedence” Some headaches with extra information being “volunteered” MX may stress BIND servers more (triggers search tree)

Basic Mode – All packets retrieved from master server, provide linked list pointers to same host

= first block, here for

second  = second block, here for third  Etc.

Population via Brute Force
Simplest Approach:


Find enough servers that recurse

DNS servers don’t move rapidly -> Scan amortizes well across many populations

Plan out who’s going to store which blocks  Populate servers with data, references to planned other servers  Easy to validate caching, though not easy to fix a broken reference

Can populate out of order – 64 populating, verifying streams

Reverse Serial Propagation
The problem: To populate the first server, you need the address of the second server. But to populate the second, you need the addr of the third. What to do? Answer: Populate backwards. The last server points nowhere. The second to last points to the last. The third to last points to the second to last.


Multi-server – last server can equal “last server group”.

Server groups may be larger than packet capacity to list servers – just include a random set

Reverse Serial Propagation
Can be quickly and statelessly deployed
Scan networks with generic recursive probe  For each incoming request seeking to service the probe, return whatever(TTL=0) and probe with an actual block request


If a block request comes back from the recurser, populate the server  If the population packet drops, the upstream should retransmit

Move back through the file after each server group fills up  Can be much slower to populate!

Large Scale DNS Scanning[0]
Modding the scanrand codebase…

Basic scanrand concept: One half spews traffic, the other sees what comes back. Little to no communication between the two halves = fast! Scanrand manages TCP; statelessness is a bit of a novelty to it. DNS is built on UDP, raw packet manipulation isn’t even necessary (it was useful, though). Reflection: Stateless operation depends on sending packets out and forgetting, only to have the necessary data returned back in the response packet
 

TCP barely reflects back 48 bits worth of data DNS reflects back whatever was in the query = thousands of bits if you want ‘em! Crypto signatures can be embedded in the reflection

Large Scale DNS Scanning[1]
Floodable DNS Queries implemented by miname

Is anyone there, recursive or not?
Many hosts that don’t support recursion will at least provide a “NXDOMAIN”, but not all will  We need a generic query that appears to almost always work PTR -> = localhost  Everyone is localhost 

Will you recurse back to me?

Large Scale DNS Scanning[2]
Returned Results
 Direct:

You requested, they responded.  Forward Lookup: You requested, they requested back to you, you responded.
 Often


not the server that you asked that asks

 Reverse

Lookup: You requested…and someone wondered who you were asking such questions

Large Scale DNS Scanning[3]
More on Reverse Auditing

Not being too secretive
 

root@bsd:~# host domain name pointer

 

L0pht’s Antisniff project – locate sniffers and IDS’s by watching reverse DNS traffic on the LAN Miname w/ Co-opted Servers: Watch sniffers and IDS’s across the entire internet

TTL=0 on PTR replies means they continually look you up when you pass by them Hard, but not impossible, to associate reverse lookup with node scanned to attract
  

Temporally associated w/ scan Multiple scans, out of order, should provide required correlation Traceroutes may help too – we know who’s looking us up

Rendering The Flood
How much data could it be?
 [root@fire

root]# ls -l dns.log -rw-r--r-1 root root 360215644 May 28 12:42 dns.log space – could use the tools we used to visualize the complexity of arbitrary data

How are we going to visualize this?
 3D

Phentropy w/ OpenQVIS

3D Viz[0]
3D Space provides a potentially valuable perspective on extremely complex datasets Volume tools have gotten more mature
 Volsuite:
 Video

Free, Open Source, Full Color, Cross Platform, Fast
games = New Absurdly High Speed Graphics Chipsets

3D Viz[1]
“Volumetric” – Texture Based
Stacks of transparent photos  Arbitrary complexity data – it all gets blurred in

Limits positional accuracy

Totally static – you can’t animate the population

“Particle” – Vertex Based
Dots! Dots everywhere!  Arbitrary positional accuracy – floating point coordinates can be zoomed into

Limits number of particles Dynamic particles can express higher dimensional data

Very easy to make dynamic

3D Viz[2]
Static: XYZ, Color  Dynamic

Shape  Direction  Color Shift  Trajectory Shift  Brightness / “Flare”

Surprising aspect of dynamics: Motion away from source point appears to bolster memory / perception of that point

3D Viz[3]
Under Development – As yet unnamed Generic 3D Plotter
 Coordinates

are streamed in, perhaps with information about how to render each particle  A more generic version of the “Spinning Cube of Potential Doom”
 Parallel

development 

Stuff = Cool More Stuff Soon