Before we start...

•Thanks to the organizers, sponsors and volonteers for making this happen in Brasil :)

•Thank you for coming.

•I'm very happy to be here !

Agenda

Introduction

Keyboard internals

Brute forcer design

Experimental results

Conclusion & bonus !

Goals, contributions :

•Demonstrate the feasability of brute force attacks on preboot authentication passwords.

•Give a pessimist estimation of the cost of password cracking on full encryption software using a generic instrumentation methodology.

•Use this metric to adapt password length policy acording with the value of the protected assets.

Juridical environment

•Cryptographic software is mostly legalized in both North and South America and Europe.

•Wikipedia : « In China, a license is still required to use cryptography. Many countries have tight restrictions on the use of cryptography. Among the more restrictive are laws in Belarus, Kazakhstan, Mongolia, Pakistan, Russia, Singapore, Tunisia, and Vietnam. »

•Users of cryptographic software must give either a copy of their keys or plain text equivalent of any text asked by authorities in case of trial, or face prison sentences in most countries.

Crypto software poor reviews

+Governments interrests

+global business communications

+terrorism blah blah

=high risk of (cryptographic ?) backdoors

&privacy threats

Technical motivations

•Even serious developpers don't test their crypto software enough, if at all (Debian SSL bug : ~32k keys).

•Vendors (in particular Truecypt) have adopted policies where they do not cover certain attacks (eg: Plain text password leakage as we presented at Defcon 0x16, or Joanna Rutowska's evilmaid attack) leaving the

More globally

•Non tech people will say :

«if it fails just go for bruteforce ».

• Sure.. but how do you do it ?

I couldn't find a public tool myself.

And then I started to wonder...

Keyboard internals

II-1) Boot sequence overview

II-2) BIOS API for user

inputs (1/2)

Interruption 0x16 invoked via functions

:

ah=0x00 , “Get keystroke” : returns the keystroke scancode in AH and its ASCII code in AL.

ah=0x01 , “Check for keystroke” : idem, but the Zero Flag is set if no keystroke is available in the Bios keyboard buffer.

II-2) BIOS API for user inputs

(2/2)

eg : lilo password reading routine :

II-3) BIOS internals for keyboard

management

II-4) BIOS keyboard buffer

Remanance... (1/3)

•Filling the BIOS keyboard buffer (with the keyboard) :

II-4) BIOS keyboard

buffer Remanence...

•Reading the BIOS keyboard buffer (using int 0x16, ah=0x00 or 0x01) :

Demo

Simulating keystrokes by

PIC programming (from real mode)

Attack scenario :

 I/O									I/O Port
									0x64
Port
0x60

Notes :

-You can get the code for this attack from the Defcon archive (the attack is called « Invisible Man »).

-For our cracking purpose, writing directly to 0x41e is way more efficient (but that was cool, right ? ;)

Demo

Retreiving passwords from

physical memory from

userland without privileges

(up to Vista SP0)

Notes

•Bitlocker's fix in Vista SP1 (replacing any character by ' ') still leaks the password length.

•This plain text password leakage vulneability is still present on many software including Lilo and Grub if you can read from arbitrary physical memory locations (typically needs root privileges).

Brute forcer design

Challenges

•Installation & initial control flow modification (BIOS Firmware, other media, MBR replacing/patching)

•Maintaining control (BP, IVT hijack, runtime patching)

Design decisions

•We want something as generic as possible, so we will avoid application specific breakpoints etc.

•The media we boot from is irrelevant (usb/cdrom/floopy..)

•Keeping control over the control flow is a bit tricky.

•Very similar to MBR virus writting (old school !! ;)

Interrupts hijacking

•Int 0x13 : we need to proxy calls to the original int 0x13, changing disk number (dl). It also allows to detect successfull decryption

•Int 0x16 : simulate keystrokes

•Int 0x10 : for performance (we don't need display)

Full attack scenario

•Boot from our code (1 sector)

•Allocate BIOS memory

•Copy the rest of our code there

•Patch the IVT (int 0x16, int 0x10, 0x13)

•Emulate int 0x19 (copy code from original MBR to 0x00:0x7c00, jump there)

jonathan@blackbox:~/h2hc$ cat BF-OS.asm |grep -v "^;"|grep [a-Z0-9]|wc -l 902 jonathan@blackbox:~/h2hc$

Demo

Bruteforcing Lilo

Demo

Bruteforcing Grub

with MD5 hash

Demo

Bruteforcing full disk encryption with TrueCrypt 6.3

Experimental results

Result #2

The cost of hashing algorithms (MD5..) is negligible in the cracking process

Result #3 : performance

Hashing algorithms : we tried 705 passwords in 30s.

Truecrypt : 10s / password (whow !)

Metrics

(assuming a hashing

algo is used)

Search space

S = sigma (i=1,length) sizeof(charset)^i

Cost

C = O (S * 3/70 * cpu_freq/(1.6GHz) *

cost_per_hour)

Amazon EC2

Cost

C ~ 3/70 * 0.085 * sigma (i=1,length)

(sizeof(charset)^length)

Cost

Exemple :

charset = [a-z] Pass length = 5

Cost ~ $45 000

Cost

Exemple :

charset = [a-z] Pass length = 8

Cost ~ $800 000 000

Cost

Exemple :

charset = [a-zA-Z0-9]

Pass length = 8

Cost ~ $800 000 000 000

Conclusions (1/2)

-Bruteforcing is physically doable for both hashing algorithms and complex symetric systems.

-Bruteforcing remains unpractical against Truecrypt so far (6 passwords / minute, recommended pass phrases of length 20).

-This methodology, while generic, is too costly to be practical against strong passwords (unless you're .gov ?).

Conclusions (2/2)

-Not using TPM like technologies allows attackers to take advantage of distributed computing, making the brute force time irrelevant.

Bonus

Random ideas dump that could not fit anywhere else in the presentation...

Having Alan Cox code your i386 real mode backdoor

(if you can't afford a trainee...)

Faxineira.asm

Joanna Rutowska's Evilmaid

attack made generic

(trojan & sniff any software's password)

Faxineira.asm

EvilMaid made generic

•Allocate BIOS memory.

•Copy yourself (1 sector) there, jump there.

•Hijack int 0x10 : save any pressed key to a 16 bytes buffer, then jump to old handler.

•Copy old MBR at 0x00:0x7c00

•Jump to 0x00:0x7c00

Bootkit/Rootkit : MBR ? floppy ? usb drive ? Cdrom ?

Network connections from

bootloaders

(without coding your own network stack)

Other possible attacks

•Timing attacks (count ticks using rdtsc)

•Glitching (won't work :-( )

•Getting physical :

FPGA (for hashing algos only : nsa@home project)

A few more things on

TrueCrypt 6.3

Truecrypt's policy and assumed attack surface

•No TPM support. Won't happen.

•No support against root or physical attacks (bootkits, trojaning ...)

•Regarding full disk encryption (the real thing why TC is great) : no keyfiles support as of version 6.3.

No TPM means

•No hardware sealing.

•We can modify the bootloader.

•We can scale on hardware/virtualisation.

Key/pass repudiation

•Setting a new key/passphrase pair is not enough : one needs to fully decrypt the drive, and then fully re encrypt it.

•Old key/pass pair would still be valid otherwise.

Forensics : HD dump vs.

Rescue iso image

•They contain exactly the same crypto information (salt+keys : only password is missing).

•We can very well brute force from a Rescue cdrom image (easier to clone/steal than a whole HD).

•This is not intuitive : social engineering risk increased.

Demo

Reversing the

Truecrypt Rescue disk

Valeu pela presenca ;)

Questions ?