# Evan A. Sultanik, Ph.D.

Evan's First Name @ Sultanik .com

Computer Security Researcher
Trail of Bits

Universitato Drexel Colegio de Komputanta kaj Informa Teknologio
Departemento de Komputscienco

## 粤式蒸鱼 (Steamed Fish with Hot Oil)

### In which I adulterate a classic Cantonese dish.

Steamed fish that is finished with a drizzle of hot oil is a classic Cantonese dish. The way I like to make it includes a few Japanese ingredients, and my method is a bit unorthodox.

### Hardware

• 12 inch frying pan with lid
• A small raised rack that will fit inside the pan that can provide a centimeter or so of clearance above the bottom of the pan; the removable rack that came with my toaster oven works perfectly
• A very small bowl
• A small saucepan (it can be tiny)
• A platter for serving that is large enough for the fish and deep enough to hold some sauce

### Software

• 2 tbsp. soy sauce
• 1 tbsp. mirin
• 1 tbsp. shaoxing cooking wine, or you can substitute an additional tbsp. of mirin
• 1 tbsp. sake
• 1/4 cup of either water, katsuo dashi stock, or water plus half of an instant dashi stock packet
• 1 fillet of a large flaky fish. Approximately 1 pound. I often use rockfish or wild striped bass. If you cannot find a large fillet, a smaller whole fish can be used.
• 1/4 cup coarsely chopped cilantro
• 2 in. knob of ginger, peeled and julienned
• 2 scallions, both whites and greens, julienned (or thinly sliced at an oblique angle, similar to the thicker Chinese “horse ear” cut)
• hot chili pepper, sliced thin at an oblique angle (optional)
• 1/8 tsp. five spice powder
• 1/4 tsp. toasted sesame oil
• 2 tbsp. vegetable oil

### Algorithm

• Mix the soy sauce, mirin, shaoxing wine, sake, and water/dashi in the frying pan
• Place the fish into the pan and marinade for 15 minutes, flipping once half way
• Meanwhile, combine the ginger, cilantro, scallion, optional chili pepper, five spice and toasted sesame oil in a small bowl and mix
• Remove the fish from the marinade, place the rack into the pan, and the fish onto the rack
• Bring to a boil
• Once boiling, cover the pan tightly, reduce the heat to simmer, and steam for 8 minutes
• Meanwhile, start heating the vegetable oil in the small saucepan
• When the fish is done steaming, remove it and the rack from the pan, placing the fish on the platter
• Increase heat to high and let the sauce reduce to the consistency of Grade A (runny) maple syrup
• Pour the sauce over the fish and put the herb mixture on top of the fish
• When the vegetable oil is very hot, spoon it over top of the herbs

## Lenticrypt: a Provably Plausibly Deniable Cryptosystem

### or, This Picture of Cats is Also a Picture of Dogs

Back in 2009, I wrote about a thought experiment on how to subvert copyright law via plausible deniability. A couple years ago I expanded on that thought experiment by proposing a seedling idea on how to accomplish it via cryptography. Since then, I’ve slowly been developing that seedling into a functioning proof-of-concept, which has culminated in the creation of the Lenticrypt project:

Lenticrypt can generate a single ciphertext file such that different plaintexts are generated depending on which key is used for decryption:

$python lenticrypt.py -e key1 plaintext1 -e key2 plaintext2 -o output.enc$ python lenticrypt.py -d key1 output.enc | diff - plaintext1 -s
Files - and plaintext1 are identical

$python lenticrypt.py -d key2 output.enc | diff - plaintext2 -s Files - and plaintext2 are identical  Unlike alternative plausibly deniable cryptosystems like the recently discontinued TrueCrypt—whose ciphertext size grows in proportion to the number of plaintexts (i.e., hidden volumes) it encrypts—Lenticrypt’s ciphertext size is proportional to the largest plaintext it encrypts. This is because Lenticrypt shares bytes in the cyphertext between each of the plaintexts it encrypts; they are not stored in separate regions of the ciphertext. Therefore, there is no straightforward way to estimate the number of plaintexts that are “hidden” inside a single ciphertext. In fact, Lenticrypt has the theoretical property that, under reasonable assumptions, there is always a near 100% probability that there exists an key in the public domain that will decrypt a given ciphertext to any desired plaintext, even if that key is not known. Therefore, even if an incriminating plaintext is revealed, the author of the ciphertext can plausibly deny having created it because there is a non-zero probability that the plaintext was legitimately decrypted by random chance. More technical details on the cryptosystem as well as additional use-cases are described in Issue 0x04 of The International Journal of PoC||GTFO. Note that Issue 0x04 of PoC||GTFO is a polyglot: Among other things, you can also treat the .pdf file as if it were a .zip. If you extract it, there are some neat Leticrypt-related Easter eggs inside the feelies. ## PoC‖GTFO Issue 0x04 ### TRACT de la SOCIÉTÉ SECRÈTE de POC ‖ GTFO sur L'ÉVANGILE DES MACHINES ÉTRANGES et autres SUJETS TECHNIQUES par le prédicateur PASTEUR MANUL LAPHROAIG ## PoC‖GTFO Issue 0x03 ### AN ADDRESS to the SECRET SOCIETY of POC ‖ GTFO concerning THE GOSPEL OF THE WEIRD MACHINES and also THE SMASHING OF IDOLS TO BITS AND BYTES by the Rt. Revd. Dr. PASTOR MANUL LAPHROAIG ## Random Solutions are Often Good Enough ### Sometimes taking the easy way out isn’t nearly as bad as it might seem! ### Hard Problems Computer Scientists and Software Engineers deal with computationally hard problems all the time. But those challenges don’t solely apply to those people creating the software; they are equally relevant to those people analyzing it. For cybersecurity researchers, finding a hash collision… determining the user inputs that can assign a certain value to a tainted EIP… deciding whether a black-box binary is malicious… they’re all really hard! As our work becomes more and more automated with the help of program analysis techniques, these computationally hard problems become more and more prevalent. But what makes one problem harder than another? In the remainder of this post I’ll answer that question, and introduce some surprising new results of my research that demonstrate how good solutions to inherently very hard problems can be quickly generated. Read on for more. ## Physical Security Followup ### These Locks are Everywhere! First of all, thanks for all of your positive feedback on our recent post on physical security. One of the comments we’ve received multiple times is that these types of locks and the practice of using mnemonics for their codes is primarily limited to government facilities. It turns out that’s not really a limitation. Like a muted post horn, now that these locks have piqued our curiosity we’re starting to see them everywhere we look. Check out this find we made today: Don’t bother checking, we scrubbed the Exif. Bank of America: Feel free to contact us directly if you’d like to know the location ;-) Another thing we neglected to emphasize in the original post was that our time estimates are upper bounds based off of the assumption that the lock will have a four minute timeout for successive failed attempts. For locks that do not have such a timeout (e.g., basically all mechanical locks, and presumably the one on this ATM), you can divide the brute-force cracking time by at least a factor of four. ## Exploiting Password Weaknesses in Physical Security ### In which I channel the spirit of my eighth academic cousin thrice removed, Richard Feynman. Digital spin locks like the Kaba Mas X-09 and X-10 are very common for high security applications like vault doors. US General Services Administration approval means that they are nearly ubiquitous in securing government filing cabinets that contain documents that are classified secret, as well as securing the doors to laboratories in which secret code is written and stored. Strolling through the hallways of any Department of Defense facility or government contractor’s office, one is likely to see many of these on the doors:  Kaba Mas X-09 These types of locks are very sophisticated: they are self-powered by rotating the dial and they have a digital printout on the top of the code as it is being entered. Security features to thwart automated brute force attack include random jumps in the numbers as the dial is being turned, algorithms for detecting “non-human” rotations of the dial, and long timeouts for successive failed attempts. These types of locks are usually configured to accept a six-digit code that is input as a sequence of three numbers from 0 to 99. That means there are$math$10^6 = 100^3 = 1,000,000$math$possible combinations. With a four minute timeout, brute forcing all of the combinations would take almost eight years! Humans are generally bad at remembering long sequences of numbers, and in most facilities the combinations to these locks are changed frequently. A single employee might have to memorize dozens of different combinations (one for each safe, door, etc.). Therefore, as a mnemonic, the security officers who set the codes often use six-letter words which are translated into codes using the mapping from a phone keypad. In fact, one will almost always see a magnet or sticker like this right next to the lock: For example, the word “HACKER” would equate to the code 42-25-37.  1 2 abc 3 def 4 ghi 5 jkl 6 mno 7 pqrs 8 tuv 9 wxyz 0 And therein lies a vulnerability. First of all, using a phone dial pad for the mapping from letters to numbers means that the numbers zero and one will never be used! This means the number of potential codes to test has been decreased from one million to$math$8^6 = 262,144$math$. That’s a decrease of almost 75%! But we can do even better. Next, let’s assume that the code word is a six-letter English word that occurs in the dictionary. Looking at /usr/share/dict/words, there are only 17,706 six-letter English words that have direct mappings using a phone keypad. That’s much lower than the 262,144 possible codes without zero and one, which means that we have reduced the number of possible codes to brute force by over 98% from the original! For example, the code 55-57-59 doesn’t appear to correspond to any words in the dictionary. That’s likely because almost all words need at least one vowel, and the phone keys 5, 7, and 9 don’t provide any vowels. Furthermore, many of those 17,706 words actually map to the same code. For example, the words “DOSAGE”, “ENRAGE”, “FORAGE”, and “FORBID” all map to the code 36-72-43. If we just count the number of distinct codes generated by those six-letter English words, we get only 14,684, a reduction of almost 99%! With a four minute failure timeout, brute forcing those 14,684 codes would only take about forty days. But wait, there’s more! Not all letters occur in English words with the same frequency; you’re much more likely to find the letter ‘e’ in a randomly chosen word than the letter ‘x’. For example, almost 10% of the six-letter English words in the dictionary end in the code “37”, just like in our example word “HACKER”. Therefore, if we cleverly order our brute force approach to try codes that occur most frequently in English words, then we can increase our chances of cracking the code quicker. I sorted all of the 14,684 distinct codes in decreasing order according to the number of words that map to that code. If a code word is chosen uniformly from the set of all six-letter English words, then this ordering will maximize our chances of cracking the code quickest. In fact, after fewer than 250 attempts, we will have cracked the code with about 5% probability. In other words, with only about half a day of access to the lock, a brute force attempt using this ordering will have a 5% chance of success. In contrast, if the lock’s code were instead set using a randomly generated six-digit number, it would require on average 50,000 attempts (~5 months) to crack it with 5% probability. On the other hand, if the code was set from a six-letter English word using a phone keypad mapping, the lock can be cracked with 50% probability after about 16 days. Codes are not usually changed on a month-to-month frequency, so an adversary with intermittent access to the lock (e.g., after hours) could quite conceivably crack it before the code were changed. If there exists an additional vulnerability that allows the attacker to detect whether each successive two-digit subcode is correct before entering the next two-digit subcode, then the expected value for the length of time required to crack the code is on the order of minutes. So what can we do about it? The most obvious solution is to simply choose codes randomly from the set of all possible one million 6-digit numbers. If we really do want to have a mnemonic, though, we could do a lot better by using a letter-to-number mapping that makes use of zeros and ones, and also produces a relatively even distribution of codes from English words. Here is one such example of a mapping:  1 a 2 e 3 br 4 co 5 dnq 6 ip 7 lmjz 8 hkt 9 su 0 fgvwxy It’s not necessarily the best one, though. Finding the optimal mapping—the one that ensures a uniform distribution across all six-digit code for six-letter English dictionary words—is actually an instance of the multiple knapsack problem (a variant known as the multiple subset sum problem) and is therefore NP-Complete, and likely intractable for problems even as “small” as this one. Let$math$L = \{a, b, \ldots, z\}$math$be the set of letters in the alphabet and let$math$f : L \rightarrow \mathbb{N}$math$be a function mapping letters to their frequency among six-letter words in the dictionary. Then we can relax the problem to the minimization of the mean squared error over the summed frequencies of the letters mapped to each number: Find a function$math$t : L \rightarrow \{0, \ldots, 9\}$math$that maps letters to numbers such that$equation$\frac{1}{10}\sum_{i=0}^9 \left(\sum_{\ell \in t^{-1}(i)} \left(f(\ell) - \frac{\sum_{\ell \in L} f(\ell)}{10}\right)\right)^2$equation$is minimized. The total number of legal mappings,$math$t$math$, is$math$10^{26} = 100$math\$ million billion billion. With that said, the example mapping above will be at least a million billion billion times more secure than using a phone keypad mapping!

The moral of the story here isn’t that these locks are inherently vulnerable; if used properly, they are in fact incredibly secure. The takeaway is that these locks are only as secure as the codes humans choose to assign to them. Using a phone keypad mapping on six-letter English dictionary words is the physical security equivalent of a website arbitrarily limiting passwords to eight characters.

## Unambiguous Encapsulation

### Defending Against “Packet in Packet” Attacks

A couple days ago, Dominic Spill and Michael Ossman presented an interesting talk at Shmoocon on using specially crafted error correcting codes to have unambiguous encapsulation, preventing attacks like “Packet in Packet.” This appears to be the culmination of some work that was first proposed to the LANGSEC mailing list back in March. The underlying problem that they are trying to address is that the physical layer is almost always error-prone: bits often get flipped. The data link layer addresses this by using error-correcting codes: methods of encoding the bits to be transmitted such that certain types of errors can be caught and corrected, at the expense of some additional bandwidth overhead.

### Isolated Complementary Binary Linear Block Codes

Spill and Ossman propose using a special error correcting code that has a property they call “isolation:” the code can be split into multiple subcodes that have additional error-correcting properties between them. One of the subcodes could be used strictly for bits of the header/preamble, while another subcode could be used strictly for the body. This provides additional error correcting capability specifically for cases like that which the attack above exploits, without increasing the bandwidth overhead of the code. They call this family of codes “Isolated Complementary Binary Linear Block Codes” (ICBLBC).

As an example, consider the following eight code words:

 00000 00011 00101 01001 01110 10110 11010 11100

Note that each code word differs from the others by at least two bits. This means that we can detect single-bit errors. For example, if 01110 is corrupted in the physical layer to 01010, then we can immediately know that an error occurred because that code does not exist in our set of eight. If two or more bit-errors occur, then an error may go undetected, which is the case that Packet in Packet attacks exploit. This is exactly like the types of error correcting codes used in most data link layer protocols. The difference, however, is that this particular set of eight code words has an additional property: it is an isolated complementary binary linear block code. Let’s call the first row of codes “Subcode 1” and the second row of codes “Subcode 2.” Now, note that every code in Subcode 1 differs from every other code in Subcode 2 by at least three bits. That means that we can not only detect but also correct single-bit errors between the two codes, giving the two subcodes properties of a Hamming Code. If we encode the header/preamble of the packet using one subcode and the body using the other, then we can circumvent Packet in Packet attacks with little to no extra bandwidth overhead!

### Generating the Codes

Spill and Ossman have provided some code that can enumerate these codes. Currently, they have reported discovery of 6-bit ICBLBCs with up to twenty-two codes in subcode 1 and two codes in subcode 2, or up to ten codes in each if they are divided evenly. While their implementation does work, it is limited by the fact that it essentially performs a brute-force search to discover them, causing them to investigate a hardware-based approach using FPGAs.

Out of curiosity, I wondered if this problem could be solved more efficiently using constraint reasoning. Therefore, we implemented a generator that encodes the ICBLBC constraints as a Satisfiability Modulo Theories problem and used DOCO (Digital Operatives’ proprietary constraint optimizer) to enumerate all feasible solution codes. Sure enough, our solver was able to almost instantaneously reproduce Spill and Ossman’s results, as well as enumerate codes of larger bit length that would likely be intractable to discover using brute force search. For example, in less than a second it was able to discover this 7-bit ICBLBC consisting of fifty codes in subcode 1 and two codes in subcode 2:

Subcode 1Subcode 2
37, 118, 44, 21, 35, 28, 107, 61, 41, 25, 94, 13, 56, 84, 124, 31, 110, 117, 50, 76, 62, 49, 47, 81, 7, 19, 109, 87, 103, 22, 93, 82, 70, 26, 127, 115, 55, 52, 59, 73, 11, 91, 88, 67, 69, 42, 121, 14, 38, 1220, 96

as well as this one with nineteen codes in each subcode:

Subcode 1Subcode 2
23, 89, 17, 27, 9, 65, 51, 5, 83, 85, 0, 18, 80, 54, 36, 33, 113, 3, 11676, 122, 61, 127, 103, 14, 107, 104, 79, 56, 109, 98, 28, 70, 94, 47, 110, 74, 42

This 8-bit ICBLBC was discovered in a matter of seconds and consists of eighty codes in one subcode:

Subcode 1Subcode 2
1, 4, 7, 8, 13, 16, 19, 25, 30, 32, 35, 37, 41, 42, 44, 47, 49, 54, 56, 59, 61, 69, 76, 81, 84, 87, 88, 97, 100, 104, 109, 110, 112, 121, 122, 124, 127, 128, 131, 133, 137, 142, 145, 146, 148, 152, 155, 161, 162, 164, 168, 171, 173, 176, 179, 181, 185, 190, 193, 196, 199, 200, 205, 208, 211, 217, 218, 220, 223, 224, 227, 229, 233, 234, 236, 239, 241, 246, 248, 25366, 75

I have reimplemented this approach using Microsoft’s “free” (as in beer) Z3 constraint solver and pushed it to a fork of Spill and Ossman’s project, available here. As always, feedback on this new approach is greatly appreciated.