Wednesday, January 25, 2017

HTC 10 Won't Connect to T-Mobile International Data Roaming

I bought an unlocked HTC 10 through a 3rd party vendor and have enjoyed the phone. I use T-Mobile service because of their great value and their free 2G international data roaming. Unfortunately, I ran into trouble connecting to cellular data the first time I traveled outside the USA with my new phone.

After arriving in Hong Kong, I received the helpful T-Mobile text message telling me I had data roaming available:

"Welcome to Hong Kong! Your T-Mobile plan gives you unlimited data at 2G speeds, calls at 20 cents/min and free texts."

Text messages and phone calls were working, but cellular data refused to connect.

Troubleshooting steps I tried:
  • Confirmed data roaming is on
  • Put phone in airplane mode for 1 minute
  • Restarted phone
  • Tried manually connecting to every carrier. Only the automatically selected one worked
  • Enabled roaming via USSD code: #766#
  • Tried manually specifying each network type (2G, 2G/3G, 2G/3G/4G)
  • Tried my T-Mobile SIM in a different phone (worked)
  • Tried a local SIM from the roaming carrier (Hutchinson 3) in my phone (worked)
  • Tried a different T-Mobile SIM, that I know works, in my phone (didn't work)
  • Asked T-Mobile if there was any block on my number for international roaming (nope)
Much of this was suggested by the very helpful T-Mobile support team and though we whittled down the problem to something specific with my phone, they never determined the exact issue. Finally, HTC customer support suggested I check my APN settings. Sure enough, they were different between my 2nd working phone and my HTC 10.

There's a known issue with the LG G3 phone not having and not automatically downloading the correct APN settings to allow T-Mobile international roaming to work. Apparently, this is also true with my HTC 10 phone, though no amount of Googling mentions it.

I believe the HTC 10 is supposed to read its APN settings from the SIM card, but instead, when a new SIM card is inserted, it prompts the user for the carrier associated with this SIM. From this user selection it populates the APNs. It appears, this hard coded list is outdated, even after an update to Android 7.0.

With the problem finally understood, several sites mention how to manually enter an APN which enables roaming. The steps are:
  1. Open Settings > Mobile Data > Access point names
  2. Copy down the settings of the currently selected Access Point
  3. Create a new Access Point with:
    1. Name = RoamingAPN
      • This can be any name
    2. APN =
    3. APN protocol = IPv4
    4. All other settings the same as the currently selected APN
  4. Save (with the triple dot menu)
  5. Select the new "RoamingAPN"
My HTC 10 connected within 15 seconds and displayed the H icon in the notification tray.  Though, you may need to reboot after making this change.

This worked for me and hopefully will help other HTC 10 owners.


Saturday, January 14, 2017

Sense Home Energy Monitor App Won't Connect

I'm very excited to learn about my personal electricity usage with my Sense Home Energy Monitor, now that it's working, but installation did not go smoothly.

The Story

I physically installed the device in my electric panel as described in the install guide:
  1. Connect the WiFi antenna external to the box
  2. Plug in the black and red wires to a 20A, 240V breaker
  3. Clip the two current sensors around my mains  
I turned on the breaker, got a happy little chime from the orange box after about 20 seconds, and saw an internal light blinking. Easy.

Unfortunately, using the app to make an initial connection is where the frustration began.

The short story is that the device refused to connect to my WiFi router, until after it upgraded its software version, which it does over the WiFi connection... Catch 22.  But figuring out this problem was painful, due to a highly simplified phone app, almost no other physical interface on the device, and very slow tech support.

First try, Android app continually fails at "Verifying Install":

Second try, iPad app continually fails at "Registering":

Both apps after failing in two different spots, with no helpful error messages, gave a page that says "Try Again".

I tried this about 15 times, power cycling the device each time, until finally it worked from the iPad.  I made no changes of any kind between the failures, except trying again. So it's still a mystery why the 15th time WiFi connected. I was able to create a log in, get the app past initialization, and then I'm given this:

The device is "Offline".

So it managed to join my WiFi network, but can't send any packets (which I confirmed from the router). More power cycling, and switching to 2 different WiFi networks, and still offline. Thankfully, a friend also has a working Sense, showed me his app, and we noticed that my device software version was older than his.

Mine: 1.0.1277-be639d4-master
His:    1.1.1353-d08ad0f-master

So guessing/hoping that "fixed in the next version" was true, I turned on tethering on my phone, and tried connecting the device to it from my iPad. This worked and he device got online. Then I left my phone sitting next to my electrical panel, checking back periodically, if it had auto-updated itself. There is no "Update Software" option in the app.  After 24 hours and 82MB of data, my device now showed version 1.1.135.  Switching back to my normal WiFi network... it still didn't work! But remembering the first rule of Windows tech support, I turned it off then back on again, and hallelujah it showed "Online" and connected to my home network.

So after 4 days of troubleshooting, my Sense is now happily measuring and learning my energy usage.  We'll see what insights it comes up with over the next few weeks.


If you're having a problem getting your Sense to initially register, or after registration connect to your WiFi network, here are some tips:
  1. Try setup with both the Android and iOS app.  Though they look the same, they obviously work a little differently with the iOS app being more mature it appears.
  2. Set up a separate WiFi network, with a different physical router. Tethering my phone worked best for me. The device likes some WiFi access points and not others. 
  3. Once connected to a temporary WiFi network, leave the device alone for 24 hours to download updates. There is no way to force an update, you just have to wait.
  4. After software update, turn it off and back on, and reconnect to your original home network.

Closing Thoughts

I contacted Sense customer support while having this issue and while they were very courteous and helpful with basic debugging steps, they were not very responsive. Support is only available over email, no phone, and it took 48 hours for my ticket to even be answered initially. I'm guessing the small team is overwhelmed with recent publicity.

Here's the start of my feature request list for the device:
  • Better/any error information in the app when something fails. "Can't connect to WiFi" would've been a lot more helpful than "Try Again".
  • There's no "Upgrade Now" button in the app. Having one would've saved me 24 hours.
  • There's no hard reset, or even factory reset option in the software. Having this *may* have helped when switching from Android to iOS setup.

Sunday, October 2, 2016

TUM CTF 2016 - totp writeup

Challenge Name: totp
Point Value: 100 Base + 100 Bonus
Description: It’s time to thank Google™ for making the Internet more secure for all of us.
Tags: web

The challenge is a website with a personal diary and user database.  We can create accounts and post messages to our diary.  A public Users page is listed, which shows the admin email address.  It's likely we need to login as an admin account and view their personal diary to find the flag.

The catch is that account logins use a Google Authenticator time-based one-time password (TOTP).  Usually, these are used in conjunction with a user selected password, but this site uses only the TOTP for login.

The Google Authenticator Wikipedia article, provided in the challenge description, explains how TOTP generates a secret key (displayed in the 2D barcode) at first creation, and that key is used along with the current time to generate 6 digit codes using HMAC every 30 seconds.  Looking at the source for the registration page, we see an example of one of these keys.

<div class="col-sm-9">
    <img src="">

    <input type="hidden" name="secret" value="6PP3WKGNATQYEV2V5ELFQPMD">
    <input type="password" id="password" placeholder="One Time Password" password="password" name="otp" required class="form-control">

Though a Google API is generating the barcode, the 120-bit (base32) secret key is being generated by the challenge site.  If we can guess what secret key was generated when the admin account was created, then we can create TOTP codes based off the current UTC time and log in as admin.  I believe this was the intent of the challenge designers, as they've conspicuously provided the time the admin account was created on the Users page.

Registered at: 2015-11-28 21:21

The key we eventually recover also gives a hint that a poor random number was used.  Without source code, there is still a lot of guessing involved in how the random numbers are generated by the site.  So while I continued to investigate that, I also setup a brute force password checker to run in the background, and that found the answer.

Brute Force

The TOTP number is 6 numeric digits, which gives 10^6 possible combinations, or 1,000,000.  Not a difficult brute force attack.  However, the correct number the server will accept changes ever 30 seconds.  But since that number is based off a SHA-1 HMAC, it has an even distribution. Each number in that 1,000,000 is just as likely to be correct for 30 seconds as any other number.  Let's say we can check 10,000 of the 1,000,000 keys within every 30 second period.  If we check the same 10,000 passwords every time (say 000 000 to 010 000) then after 50 attempts (1,000,000 / 10,000 / 2) we should have guessed the right password.  That's 50 attempts * 30 seconds, or 25 minutes, which is not long at all for an online brute force attack.

We could script this, but Burp Intruder has most of the tools we need already.  First, I benchmarked how fast Burp could send login requests from my connection to the web server.  I settled on 25 threads, which on my Internet connection, over 30 seconds could make about 3000 attempts.  Meaning an attack time closer to 84 minutes.  Next, it's not easy to script Burp to restart an Intruder attack every 30 seconds, exactly on the :00 and :30 marks.  So, I went the even lazier approach of trying all 1,000,000 passwords in a row.  But since the password the server picks is an even distribution, this shouldn't hurt our attack time too much in practice.

Lastly, playing with the site while using Burp Proxy showed no difference in the response from the "POST /?target=login.php" message on an unsuccessful login vs a successful login. Instead, the PHPSESSID returned in the cookie, has server side state, which marks the login as successful.  So we'd need to do an extra "GET /" request after every POST to check for success.  Well that would slow down the attack another 2x, so instead we'll check manually every 30 seconds or so.

So the simplified steps for the brute force attack are:

1) Generate a file of every possible TOTP possibility:

$ for i in {0..999999}; do printf "%06d\n" $i >> totp_999999.txt; done

2) Start Burp Intruder with POST login requests against the web server using the generated file:

3) Periodically refresh your web browser using the same PHPSESSID cookie as Burp:

I got lucky after about 75,000 request / 13 minutes.

4) Collect the flag:

Sunday, August 7, 2016

OpenCTF 2016 - diskomatic writeup

Challenge Name: diskomatic
Point Value: 200
Description: It slices, it dices, it juliennes fries!
Tags: forensics

The challenge is a binary file with a forensics tags, so I assumed it would be pulling some data out of some other file format.

Step 1:  Unpack


$ file diskomatic-98757b0043b8aea77e9014cff8ead7b1
diskomatic-98757b0043b8aea77e9014cff8ead7b1: gzip compressed data, last modified: Mon Jun 22 00:08:37 2015, max compression, from Unix

Simple enough.

$ zcat diskomatic-98757b0043b8aea77e9014cff8ead7b1 > diskomatic.dat

$ file diskomatic.dat diskomatic.dat: DOS/MBR boot sector, code offset 0x58+2, OEM-ID "mkdosfs", FAT  1, Media descriptor 0xf8, sectors/track 32, heads 64, sectors 250368 (volumes > 32 MB) , FAT (32 bit), sectors/FAT 1941, serial number 0x20550574, label: "DISKOMATIC "


Step 2:  Mount Filesystem


So we have a FAT filesystem image.  Let's try mounting it and seeing what's inside.

$ sudo mount -t vfat ./diskomatic.dat ~/mnt/dos/
$ ls -la ~/mnt/dos
total 5
drwxr-xr-x 2 root  root   512 Dec 31  1969 .
drwxr-xr-x 3 grind grind 4096 Aug  6 12:41 ..
$ df -h | grep dos
/dev/loop0      122M   512  122M   1% /home/grind/mnt/dos
$ sudo umount ~/mnt/dos

Interesting so from the filesystem's perspective there are no valid files, but we have 122MB of something.


Step 3: Analyze the Binary


$ binwalk -e diskomatic.dat

17787904      0x10F6C00       PNG image, 1600 x 900, 8-bit/color RGB, non-interlaced
34088942      0x20827EE       Minix filesystem, V1, little endian, 30 char names, 23429 zones

binwalk found a PNG file, but was unable to extract it.  Interesting... perhaps it is corrupted somehow. I don't know where the Minix FS detection came from, but it's a false positive.

So let's open up the file in a hex editor and take a look at location 0x10F6C00.


Step 4:  Hex Editor

The beginning of the file is as `file` advertised, a FAT32 partition table.  Let's go to location 0x10F6C00.

Here's the PNG header, with the first IDAT section but it is noticeably quite short.  Searching for IDAT we find many more of these sections.  Each one is exactly 512B, which I believe is the block size for FAT32.  So it seems we have a PNG file which has been fragmented and we need to piece it back together.  Well let's try the naieve approach, and just concatenate all the chunks.  We can do this from within the hex editor by cropping the file from location 0x10F6C00 onward and then doing a find and replace on all the \x00\x00\x00\x00 words with nothing.

We'll save this file as diskomatic.png and open it.

Well that didn't work. But the CRC error is a great hint.  PNG is a simple format with chunks split into just 4 fields.

Length Chunk type Chunk data CRC
4 bytes 4 bytes Length bytes 4 bytes

But you don't have to take my word for it, you can read the Wikipedia article.

Looking at the 512B blocks in the hex editor again, we see that the PNG chunks are split in the middle of the data section, leaving some data and the CRC for that data at the front of the next block.  So we have a fragmented file, but we have a method of determining in which order the pieces are supposed to fit back together by combining two blocks and then checking whether the CRC for the PNG chunk is correct.  So now we have an idea of the problem and a solution, but first we'll need to extract each block of the PNG file.


Step 5: Programmatically Extracting FAT32 Blocks


At first I just selected a few blocks in the hex editor and saved them as new files.  But after about 6 of these, with the location not moving much, I realized I'd probably need to automate this.

$ grep -abc IDAT diskomatic.dat

Yep, I definitely don't want to copy'n'paste 5000 times by hand.  So I wrote some Python:
FILE = "diskomatic.dat"

def get_parts():
    fh = open(FILE, 'rb')

    parts = []

    # we know the location of the first part from `binwalk`
    p_idx = 0
    part = ""

        while True:
            block =
            if len(block) == 0:
            if block[0:4] != "\x00\x00\x00\x00":
                fh2 = open("parts/p"+str(p_idx)+".png.part", "wb")
                p_idx += 1
            offset = fh.tell()
            if offset % 0x100000 == 0:
                print "0x%x" % offset
    except EOFError:

This writes all the blocks/parts out, 1 per file, and puts them in a directory called parts. After running this code we have 6238 files.

$ cd parts
$ ls | wc
   6238   56144  354468


Step 6:  Reassemble the Blocks


First I tested my hypothesis that all the blocks we needed were present, by programmatically concatenating every block with the first block, and was successful finding one block.

From that code, I moved on to checking every block against every other block, and constructing the correct ordering of blocks in a dictionary.  Then writing that string of blocks out into a single PNG file.  The full code, excluding the get_parts() function is:
#!/usr/bin/env python

import binascii
import struct

FILE = "diskomatic.dat"

def tokenize_chunk(data):
    length = struct.unpack("!I", data[0:4])[0]
    chunk = data[0:4+4+length+4]

    return chunk

def png_chunk_verify_crc(chunk):
    length = struct.unpack("!I", chunk[0:4])[0]
    ctype = chunk[4:8]
    data = chunk[8:8+length]
    crc = struct.unpack("!I", chunk[8+length:8+length+4])[0]

    # The CRC is a network-byte-order CRC-32 computed over the chunk type and
    # chunk data, but not the length.
    crc_c = (binascii.crc32(chunk[4:-4]) &amp; 0xFFFFFFFF)

    valid = (crc == crc_c)

    return valid

def read_parts():
    parts = []

    # XXX: magic number: that's how many chunks were in the original file
    for i in xrange(0, NUM_CHUNKS):
        fh = open("parts/p"+str(i)+".png.part", "rb")
        data = bytearray(

    return parts

def compare_parts(parts):
    parts_dict = {}

    for i in xrange(1, NUM_CHUNKS):
        first_two = parts[0] + parts[i]
        # XXX: magic number, for first one, because there's the PNG header
        chunk = tokenize_chunk(first_two[0x2E:])
        matched = png_chunk_verify_crc(chunk)
        if matched:
            print "Chunk 0 and %d matched!" % i
            parts_dict[0] = i

    for i in xrange(1, NUM_CHUNKS):
        matched = False
        for j in xrange(1, NUM_CHUNKS):
            if i == j:
            two_parts = parts[i] + parts[j]
            # XXX: magic number, most other parts seem to have the same offset
            offset_to_chunk = 0x2E
            chunk = tokenize_chunk(two_parts[offset_to_chunk:]) 
            matched = png_chunk_verify_crc(chunk)
            if matched:
                if i % 100 == 0:
                    print "Chunk %d and %d matched!" % (i, j)
                parts_dict[i] = j

        if not matched:
            print "Chunk %d had no match!" % (i)

    return parts_dict

def write_parts(parts, parts_dict):
    fh = open("final.png", "wb")

    next_key = 0
    for i in parts_dict.iterkeys():
            next_key = parts_dict[next_key]
        except KeyError:
            print "Hit KeyError"


def main():
    parts = read_parts()
    parts_dict = compare_parts(parts)
    write_parts(parts, parts_dict)

if __name__ == "__main__":

The nested for loops do some unnecessary work, but the whole thing runs in under 5 minutes on my laptop.


Step 7:  Collect the Flag


Now we have a (hopefully) valid PNG. Let's open it.

Friday, June 3, 2016

Running IDA Evaluation Version on Kali Linux 2016.1 64-bit

The IDA Evaluation Version previously came installed with Kali Linux 1.0, but since the upgrade to 2.0 and now Rolling Edition, IDA is no longer present.  Since the evaluation version is available as 32-bit binaries only, getting it running requires figuring out the rather large set of dependent 32-bit libraries that must be installed on Kali 64-bit.

For a quick fix, run the following commands:

$ sudo dpkg --add-architecture i386
$ sudo apt-get update
$ sudo apt-get install libglib2.0-0:i386 libx11-xcb1:i386 libxi6:i386 libsm6:i386 libfontconfig1:i386 libqt5gui5:i386

Now IDA should successfully execute from the CLI and give you a graphical window to accept the IDA License Agreement.

~/Downloads/idademo69$ ./idaq

How to determine what libraries are missing and which packages provide them involves the iterative process of:
  1. Check for missing shared objects
  2. Check which package provides them
  3. Install that package
For example:

~/Downloads/idademo69$ ldd idaq | grep "not found" => not found => not found => not found => not found => not found => not found => not found
~/Downloads/idademo69$ dpkg -S
libxext6:amd64: /usr/lib/x86_64-linux-gnu/
libxext6:amd64: /usr/lib/x86_64-linux-gnu/
~/Downloads/idademo69$ sudo apt-get install libxext6:i386



Saturday, March 28, 2015

32-bit to 64-bit Conversion Woes with strtoul()

Proper error checking when using the strtol() family of functions is notoriously difficult.  See this stackoverflow thread explaining it.

Focusing only on strtoul() here, the correct error checking in most cases is:

unsigned long
mystrtoul(char *str, bool *success)
    unsigned long val = 0;
    char *eptr = NULL;
    errno = 0;

    *success = true;

    val = strtoul(str, &eptr, 0);
    if (eptr == str || *eptr != '\0' ||
        (val == ULONG_MAX && errno == ERANGE) ||
        (val == 0 && errno == EINVAL))
        *success = false;

    return val;

That is:
  • IF your end pointer points to the beginning of your string, you've got a problem
  • IF your end pointer doesn't point to the end of your string, you've got a problem
  • IF your output is at the max end of the range (a potentially valid value) AND you get errno, you've got a problem
  • IF your output is 0 (a potentially valid value) AND you get errno, you've got a problem
  • ELSE congratulations, you've converted a string to an integer
Understandably, a lot of people get this error checking wrong when using this function. I've reviewed source code I work on and found mistakes.  But even if you got all of this right, you may still have gotten it wrong.

If you originally wrote this code for a 32-bit system, with the intention of converting and storing a 32-bit number, then what you wrote is correct.  If you then recompile this correct code on a 64-bit system, it becomes incorrect.

That's right with no modifications (and no compiler warnings), your carefully reviewed integer conversion function goes from right to (dangerously) wrong.

For example give as input "4294967296" and assign to an unsigned int:

    bool success = false; 
    unsigned int val = 0;

    val = mystrtoul("4294967296", &success);
    printf("%s %u\n", success?"success":"failure", val);

    return 0;

  • On 32-bit Linux:  failure 4294967296 (ERANGE)
  • On 64-bit Linux:  success 0
The 32-bit compilation catches the integer overflow and errors out.  The 64-bit version does not and rolls back around to 0.  So your output is wrong and the usual possible errors and security holes with integer overflows apply:
  • malloc(4294967296) allocate 0 bytes and a buffer overflow is created
  • uid_t 4294967296 become 0 and a user is root
And strings passed in to strtoul() are very often user controlled input.  That's why you need to convert them from an on-the-wire, CLI, or on-disk string format to an in-memory integer.

This isn't strictly a strtoul() problem.  This is a LP64 data model problem.  Any assumption about the 32-bit size of a long needs to be checked when converting from 32-bit to 64-bit runtime.  This is just a particularly tricky example of the trade offs made by LP64.

Saturday, February 4, 2012

Windows Update Error Code 0x80073712

I have several Windows 7 VMs which I use for various dev testing.  For the past 9 months all of them having been failing during Windows Update.  Most updates apply, but Service Pack 1 consistently errors during installation with the unknown error code 0x80073712.

Microsoft has a well written KB explaining steps you can take when this issue occurs:

I went through each step, running the checksur.exe and sfc.exe tools and downloading the stand alone Service Pack 1 installer.  All failed.  Next, I tried step 4 attempting to upgrade with the original Windows 7 CD.  Note, that the CD image must be the exact same used to install the original OS.  Same release flavor, Enterprise, in my case, and same SP level.  I thought I'd want to upgrade using the ISO image with SP1 built in, but no that refuses to even attempt an upgrade.

Finally, with the correct ISO image I got the clue that no previous error message had given:

  • There is not enough free space on partition (C:). A total of 16372 megabytes (MB) of free disk space is required. Try Disk Cleanup, or moving files to an external location such as a CD, DVD, or external hard drive, and then restart the installation.

I didn't have enough free space on the drive to install the Service Pack.  At first I was surprised because I had 10GB free, but in fact Win7 SP1 requires about 16GB to install.  Next, I was annoyed that this simple error wasn't raised by Windows Update in the first place.

I keep my drive space pretty sparse on my test VMs because they don't have many applications or data stored on them.  But a 40GB primary drive size is not enough for Windows 7.  I expanded it to 80GB using VMWare Workstation 8's built-in "Expand Disk" tools then used sysresccd and gparted to expand the NTFS partition across the whole disk.

Success!  Windows Update installs Service Pack 1 with no problems.  I hope this helps you or someone else avoid a half-day of IT induced profanity.