Why WEP Is Not Secure

WEP was included in the original IEEE 802.11 standard in 1997, but it was not until 1999 that systems were widely deployed when IEEE 802.11b and Wi-Fi became established. Most vendors added key extensions allowing a 104-bit key to be used (128 bit if you count the IV), and this was also adopted by Wi-Fi.

The industry started to have concerns about WEP security as designs were made and engineers started to point out some problems. In particular, the weakness of the authentication method was noted, as described earlier, and the authentication phase was dropped altogether. However, the manufacturers' concerns related to the strength of the security rather than the overall integrity. In other words, they were concerned that a serious and major attack might succeed. Nobody thought that it would be easy to break WEP.

Because of these concerns and also because of the difficulties in key management discussed in this chapter, the IEEE 802.11 Standards Committee launched a new task group to look at upgrading the security approach. The IEEE 802.11 committee is, essentially, a voluntary organization. Meetings are held roughly six times a year and the meetings are open to any industry members. The process involves technical presentations, discussions, drafting text, and a lot of voting. It is a democratic and parliamentary process. As a result, new standards do not develop quickly. Major modifications such as the security subsystem take years rather than months.

While the IEEE 802.11 group was considering a new approach, the security research community was also looking at WEP. In the early days, security gurus had not paid much attention because wireless LAN was not a widespread technology. But by 2000 it was appearing everywhere and many universities had installations. Security researchers had direct contact with Wi-Fi LAN. As did most people, they realized the potential advantages; but, unlike most people, they immediately questioned whether the security provisions were bomb-proof.

The answer came in 2000 as a series of reports emerged highlighting weakness after weakness (Walker, 2000; Arbaugh et al., 2001; Borisov et al., 2001). Finally, an attack was published showing that the keys could be extracted in a busy network within hours, regardless of the key length. The press were ecstatic. Security weakness and security breaches sell copy, and headlines appeared around the world. Even the main public media channels covered the event. This was a measure of how far IEEE 802.11 had come that it was considered of interest to the general public. However, there were lots of red faces?embarrassed manufacturers and angry customers.

This section looks in detail at the attacks on WEP that have been identified. WEP is a general term that covers a number of security mechanisms. As a basis for evaluating WEP, let's review the mechanisms that are needed for security:

  • Authentication

  • Access control

  • Replay prevention

  • Message modification detection

  • Message privacy

  • Key protection

Unfortunately, WEP fails to perform in all these areas. We'll look at each one separately.


Authentication is about one party proving to the other that he really is who he claims to be. Authentication is not a one-time process?in other words, it is not enough to prove once that you are authentic. It is only useful if you can prove it every time you communicate. The process of authentication is often time consuming, and so a common approach is to perform full authentication on first contact and then provide a limited-life "identity badge." Ideally, the identity badge is such that it cannot be transferred to someone else. A photo ID is an example in which the government or corporation authenticates you.

In the wireless world, you usually need mutual authentication. The network wants proof about the user, but the user also wants proof that the network really is the expected one. This is important for Wi-Fi LANs because it is so inexpensive to set up decoy access points.

Finally, security experts point out that it is essential to use different secret keys for authentication than you use for encryption.[3] The use of derived keys is recommended because master keys should rarely or never be exposed directly to attack. In summary, the basic requirements for authentication in wireless LANs are:

[3] It is acceptable to cryptographically derive two separate keys from a single master key. If done correctly, the two keys are effectively independent.

  1. Robust method of proving identity that cannot be spoofed

  2. Method of preserving identity over subsequent transactions that cannot be transferred

  3. Mutual authentication

  4. Independent keys independent from encryption (and other) keys

Unfortunately, WEP fails on all counts. As a reminder, WEP authentication relies on a challenge?response mechanism. First, the AP sends a random string of numbers. Second, the mobile device encrypts the string and sends it back. Third, the AP decrypts the string and compares to the original string. It can then choose to accept the device and send a success message.

The key used for this process is the same WEP key used for encryption, thus breaking rule 4. The operation does not authenticate the access point to the mobile device because a rogue access point can pretend it was able to check the encrypted string and send a success message without ever knowing the key. Hence rule 3 is broken.

Rule 2 is broken because there is no token provided to validate subsequent transactions, making the whole authentication process rather futile.

Rule 1 is rather irrelevant given the weaknesses already pointed out, but it's quite interesting to look at why this also fails.

During authentication the access point sends a random string of 128 bytes. The way in which this "random" string is generated is not defined, but one would hope at least that it was different for each authentication attempt. The mobile station encrypts the string and sends it back. Sounds good, but hang on a moment?WEP encryption involves generating a sequence of pseudorandom bytes called the key stream and XORing it with the plaintext. So any one watching this transaction now has the plaintext challenge and the encrypted response. Therefore, simply by XORing the two together, the enemy has a copy of the RC4 random bytes. Remember the basic equation:

P  R = C    (Plaintext XOR Randombytes = Ciphertext)

And remember that XORing twice gets you back to the original value (that's decryption):

If P  R = C then C  R = P

By the same argument, XORing the ciphertext with the plaintext gives you the random key stream:

If P  R = C then C  P = R

Whoops, the game's over! The attacker now knows the key stream corresponding to a given IV value. Now the attacker simply requests authentication, waits for the challenge text, XORs with the previously captured key stream, and returns the result with the previously captured IV.

To check the result, the access point appends the IV (chosen by the attacker) to the secret key and generates the RC4 random key stream. These will, of course, be the same bytes that the attacker worked out because the key and IV are the same as last time. Therefore when the access point decrypts the message by XORing with the RC4 key stream, surprise, surprise, it matches. The attacker is "authenticated" without ever knowing the secret key. Hopeless!

Although an attacker can get authenticated in this way, she can't then communicate because frames are encrypted with WEP. Therefore, she needs to break WEP encryption as well. However, there is even more bad news. For some of the methods of attacking encryption keys, the enemy needs a sample of matching plaintext and ciphertext. Sometimes this can be quite hard for an attacker to get, and there are various tricky methods that she might use to try to get such a sample. What a gift! The WEP authentication method provides a 128-byte sample free of charge. Worse, it is a sample of the first 128 bytes of the key stream, which is the most vulnerable to attack.[4] So not only does this approach not authenticate, it actually assists the enemy to attack the encryption keys. Hmmm… we better move on. Most systems today don't use the futile WEP authentication phase anyway.

[4] We will look at this vulnerability in more detail in Chapter 11 on TKIP.

Access Control

Access control is, rather obviously, the process of allowing or denying a mobile device to communicate with the network. It is often confused with authentication. All that authentication does is to establish who you are; it does not follow that, because you are authenticated, you should be allowed access.

In general, access is usually controlled by having a list of allowed devices. It may also be done by allowing access to anyone who can prove he has possession of a certificate or some other electronic pass.

IEEE 802.11 does not define how access control is implemented. However, identification of devices is only done by MAC address, so there is an implication that a list of acceptable MAC addresses exists somewhere. Many systems implement a simple scheme whereby a list of allowed MAC addresses can be entered into the access point, even when you are operating without WEP. However, given the ease with which MAC addresses can be forged, this cannot be considered as a serious security mechanism.

If you can't trust the MAC address, the only thing left to WEP is the encryption key. If the mobile station doesn't know the correct WEP key, then the frames it sends will produce an ICV error when decrypted. Therefore, the frames will be discarded and, effectively, the device is denied access. This last line of defense is really all that the original IEEE 802.11 standard has to offer.

Replay Prevention

Let's suppose you are an attacker with a wireless sniffer that is able to capture all the frames sent between an access point and a mobile device. You observe that a new user has turned on her laptop and connected to the network. Maybe the first thing that happens is that the server sends her a login message and she enters her user name and password. Of course, you can't see the actual messages because they are encrypted. However, you might be able to guess what's going on, based on the length of the messages.

Later on, you notice the user has shut down and gone home. So now is your chance. Bring up your own client using her MAC address and connect to the network. As we have seen earlier, that part is easy. Now, if you are lucky, you'll receive a message to log in. Again, you won't be able to decode it, but you can guess what it is from the size. So now you send a copy of the message the legitimate user sent at that point. You are replaying an old message without needing to know the contents. If there were no replay protection, the access point would correctly decode the message; after all, it was originally encrypted with a valid key before you recorded it. The access point passes the message to the login server, which accepts it as a valid login. You, as an attacker, just successfully logged into the network and the server. It's not clear where you would go from there. However, from a security standpoint, this is a serious breach.

There are many other examples in which a replay attack can breach security unless the network is designed specifically to detect and reject old copies of messages. The wireless security protocol should allow only one copy of a message to be accepted. Ever.

By this time, it may come as no surprise to discover that WEP has no protection against replay at all. It was just not considered in the design. There is a sequence number in the MAC frame that must increase monotonically. However, it is not included in the WEP protection so it is easy to modify the sequence number to be valid without messing with the encrypted portion on the frame.

Replay protection is not broken in WEP; it simply doesn't exist.

Message Modification Detection

WEP has a mechanism that is designed to prevent message modification. Message modification can be used in subtle ways. The first thing people think about when message modification is proposed is to change the contents of the message in an obvious way. For example, changing the destination bank account number on a deposit or changing the amount transferred. However, in reality such large-scale modifications would be very hard to mount and assume that you can read the original message and effectively forge new messages.

If you are unable to decrypt the message, it is not obvious how modifying the ciphertext would be useful. However, even in this case modifications can be used to extract information. A technical paper by Borisov et al. (2001) proposed a method to exploit "bit flipping" in which a few bits of the ciphertext are changed at a time. They pointed out that the position of the IP header is usually known after encryption because WEP does not rearrange the byte positions. Because the IP header has a checksum, changing one bit in the header causes the checksum to fail. However, if you also change bits in the checksum, you might get a match. The researchers showed that, by flipping a few bits at a time and seeing whether the IP frame was accepted, based on whether responses came back, they could eventually decode portions of a frame.

To prevent tampering, WEP includes a checkfield called the integrity check value (ICV). We looked at this briefly in the previous section on frame formats. The idea behind the ICV is simple: compute a check value or CRC (cyclic redundancy check) over all the data to be encrypted, append the check value to the end of the data, and then encrypt the whole lot. If someone changes a bit in the ciphertext, the decrypted data will not have the same check word and the modification will be detected. The thinking is that, because the ICV is encrypted, you cannot go back and correct its value to compensate for the other changes you have made. It is only intended to provide protection to the ciphertext. If an attacker already knows the keys, he can modify the data and recompute the ICV before re-encrypting and forwarding the frame. So use of the ICV protects the ciphertext from tampering?right?

Wrong again! Borisov et al. pointed out a flaw in the logic. The CRC method used to compute the ICV is called a linear method. It turns out that, with a linear method, you can predict which bits in the ICV will be changed if you change a single bit in the message. The ICV is 32 bits. Let's suppose the message is 8,000 bits (1,000 bytes) and you flip bit position 5244. You can then compute which bits in the ICV will be changed as a result. It is typically not a single bit but a combination of bits that will change. Note that we used the word "change," not "set" or "clear." You don't need to know the actual value of the plaintext; you just need to know that if you flip the value of a certain bit in the data, you can keep the ICV valid by also flipping a certain combination of its bits. Unfortunately, because WEP works by XORing the data to get the ciphertext, bit flipping survives the encryption process. Flipping a bit in the plaintext always flips the same bit in the ciphertext, and vice versa.

If you've hung in through the argument in the last paragraph, you will see that, because of the fact that bit flipping survives encryption; the assumption that the ICV is protected by encryption just doesn't hold water. Its actual value is protected, but you can reliably flip its bits. And because you can work out which bits to flip corresponding to a change in the data, you can completely defeat its protection.

With a muffled thump, another of WEP's protections just hit the floor. The reality is that WEP provides no effective protection against ciphertext modification.[5]

[5] In fact, in the general case, no integrity check word can be used successfully with RC4 unless it is created using a key generated specifically for integrity checking (as distinct from the encryption key).

Message Privacy

This is the big one: attacking the encryption method of WEP. We have seen that the other protections have already been stripped away; but, at the end of the day, if the encryption method holds up, then the attacker is very limited in what he can do. So far, it's just watching shadows or throwing rocks at the window; but if the encryption can be breached, the attacker is inside the house.

There are two main objectives in attacking the encryption: decode a message or get the keys. The ultimate success is to get the keys. Once an attacker has the keys, he is free to explore and look for the valuables. Possession of the keys doesn't automatically mean access to confidential information because there are other layers of security inside, such as server passwords and operating system protections. However, the issue of network access is put aside. Furthermore, if an attacker can get the keys, he can probably go undetected, which is important to buy the time to find useful information. If an attack is detected, the WEP keys can be changed, putting the attacker back to square one.

The next best thing to getting the keys is to be able to get the plaintext. If you can get the plaintext in a reasonably fast and reliable way, you have access to a range of other types of attacks using message modification and replay. That information can also be used as a stepping-stone to getting the keys.

There are three weaknesses in the way RC4 is used in WEP and we will look at each case separately:

  • IV reuse

  • RC4 weak keys

  • Direct key attack

IV Reuse

One of the first cryptographers to point out weaknesses in WEP was Jesse Walker of Intel. In October 2000 he wrote a submission to the IEEE 802.11 Standards Committee entitled "Unsafe at any key size: An analysis of the WEP encapsulation." This title was designed to get attention?and it did. Walker pointed out a number of potential weaknesses, but especially focused on the issue of IV reuse.

Let's quickly review how the IV is used. Instead of using a fixed secret key, the secret key is appended to a 24-bit IV value and then the combined IV/ secret is used as the encryption key. The value of the IV is sent in the frame so the receiving device can perform the decryption. One purpose of the IV is to ensure that two identical messages don't produce the same ciphertext. However, there is a second and more important purpose related to the way WEP uses XOR to create the ciphertext.

Let's suppose for a moment that there was no IV and only the secret key is used for encryption. For every frame, the RC4 algorithm is initialized with the key value prior to the start of the pseudorandom key stream generation. But if the key were to remain fixed, the RC4 algorithm would be initialized to the same state every time. Therefore the key stream produced would be the same sequence of bytes for every frame. This is disastrous because, if the attacker can figure out what that key stream is, he can decode every frame simply by XORing the frame with the known sequence. He doesn't need to know the key.

By adding the IV value to the key each time, RC4 is initialized to a different state for every frame and so the key stream is different for each encryption?much better. Let's review that statement because there is an implicit assumption: The IV value is different for every frame. If the IV is a constant value, you are no better off than in the static key case.

So we see that constant IV is useless. We can also see that using a different IV for every frame, and I mean every frame, is a good idea. What about the middle ground? There are a limited number of possible IVs so it is acceptable to use a different IV for most frames but eventually start reusing IVs that have been used in the past. The simple answer is that this is not acceptable?but it is precisely what WEP does.

Let's look at why IV reuse is a problem. We have said that the IV should be different for every frame. However, the original IEEE 802.11 standard did not say how it should be generated (actually it did not require that it be changed at all!). Intuitively you might think that the best approach would be to generate a random value. However, with random selection there is a good chance that you will get a repeating IV quite quickly. This is known as the birthday paradox (see sidebar).

In the case of IVs, it means that you are likely to get a duplicate IV sooner than you expect if you pick random values.

The best way to allocate IVs is simply to increment the value by 1 each time. This gives you the longest possible time before a repeating value. However, with a 24-bit IV, an IV collision (use of a previous value) is guaranteed after 224 frames have been transmitted (nearly 17 million). IEEE 802.11b is capable of transmitting 500 full-length frames a second and many more shorter frames. At 500 frames a second, the IV space would be all used up in around seven hours.

In reality a collision is likely much sooner because there may be many devices transmitting, each incrementing a separate IV value and using it with the same key (assuming default keys are in use). Implementation errors can compound the problem. At least one major Wi-Fi LAN manufacturer always initializes the IV counter to 0 when the system is started up. Imagine that ten users come into work and start up their laptops. Depending on who does what, the IV counter of some will get ahead of others, but there will be a rich harvest of IV collisions to be had by an observer. IV collisions are a fact of life for WEP so let's look again at why collisions are a problem.

The Birthday Paradox

When you meet someone, there is only a 1 in 365 chance that the person has the same birthday as you. However, the chance of meeting someone with your birthday increases surprisingly fast as you meet more people. In fact, there is a 50% chance that you will find someone with a matching birthday within the first 25 people you meet. This is a surprising fact, which is probably why it is called a paradox.

If you know the key stream corresponding to a given IV value, you can immediately decode any subsequent frame that uses the same IV (and secret key). This is true regardless of whether the secret key is 40 bits, 104 bits, or 1,040 bits. To decode every message, you would have to know the key stream for every possible IV value. Because there are nearly 17 million possible IV values, that seems like a daunting task. However, it's not impossible: If you want to store a 1,500-byte key stream, for every possible IV you need a storage space of 23Gbytes?quite feasible on the hard disk of an everyday computer. With such a database, you could decode every message sent without ever knowing the secret key. However, you still have to find out all those key streams and that's not so easy.

Suppose you have captured two messages encrypted using the same IV and secret key. You know that the key stream is the same in both cases, although you don't know what it is yet. Using our simple notation:

C1 = P1  KS       (Ciphertext msg1 = Plaintext msg1 XORed Keystream)


C2 = P2  KS       (KS is the same in each case)

If you XOR C1 and C2, KS disappears:

C1  C2 = (P1  KS)  (P2  KS) = P1  P2  KS  KS = P1  P2

This is true because XORing the same value twice takes you back to your original value.

So the attacker now has a message that is the XOR of two plaintexts. Is that useful? No not yet. However, some of the values of plaintext are definitely known, such as certain fields in the header. In other fields the value is not known, but the purpose is known. For example, the IP address fields have a limited set of possible values in most networks. The body portion of the text often encodes ASCII text, again giving some possible clues.

Over a period of time, if you collect enough samples of duplicated IVs, you can probably guess substantial portions of the key stream and hence decode more and more. It's like a collapsing building: Each block you knock away makes it more likely that the whole lot will fall down. It's hacker's celebrity squares in which you have some of the letters in a word and you try to guess the whole sentence. But once you succeed for a given IV, you can decode every subsequent frame using that IV and generate forged frames using the same IV. All without knowing the key.

This characteristic of WEP was worrisome and resulted in IEEE 802.11 undertaking the new security design. However, it was not considered a major threat to everyday use. After all, it would take a huge effort to decode a significant number of frames and the need for intelligence in guessing the plaintext makes it hard to create an automatic script tool. So the cryptographers cringed and the manufacturers worried, but the world went on after this attack was publicized. But there was worse to come.

RC4 Weak Keys

The fundamental part of RC4 is not encryption but pseudorandom number generation. Once we have a string of pseudorandom bytes, we can use them to encrypt the data by the XOR function. As we have seen, using this simple XOR function is a source of weakness if it is not applied correctly; but for the moment, let's concentrate on the pseudorandom sequence, or key stream.

RC4 works by setting up a 256-byte array containing all the values from 0 to 255. That is, each value from 0 to 255 appears once and only once in the array. However, the order in which they appear is "randomized." This is known as a permutation of the values. What's more, the values are reorganized continuously as each pseudorandom byte is generated so there is a different permutation of the array each time.

Each pseudorandom byte is generated by picking a single value from the permutation based on two index values, i and j, which also change each time. There are very many permutations (or arrangements) of 255 values that can be made. In fact, combined with the two indices, there are 512 * 256! (factorial) possibilities, a number too big to compute on any calculator we have, even using scientific notation.

This property of RC4 makes it very powerful despite its simple implementation. It is amazingly hard to distinguish an RC4 pseudorandom sequence from a real random sequence. RC4 has been studied by many cryptographers and yet the best known method for distinguishing an RC4 stream from true random data requires a continuous sample of 1Gbyte of the stream before it can reliably decide that the stream was generated by RC4. For WEP, of course, we already know RC4 is used, but this fact gives you some idea how effective RC4 really is once it gets going.

The phrase "once it gets going" in the last sentence is important. It signals the fact that RC4 has a potential weakness. To understand the weakness, let's quickly review how RC4 works. First it creates a table (the S-box) with all the values 0?255. Then it creates a second 256-byte table with the key, repeated over and over until the table is full. Then it rearranges the S-box based on values in the key table. This is the initialization phase. The first pseudorandom byte is generated by rearranging the S-box again and picking a byte.

The problem here is that there are not many rearrangements between the initial setup of the key table and the first pseudorandom byte. Fluhrer et al. (2001) analyzed this fact, resulting in their now famous paper "Weaknesses in the Key Scheduling Algorithm of RC4." They showed that for certain key values, which they called weak keys, a disproportionate number of bits in the first few bytes of the key stream (pseudorandom bytes) were determined by a few bits in the key itself.

Let's look at this weakness in another way: Ideally if you change any one bit in the key, then the output key stream should be totally different. Each bit should have a 50% chance of being different from the previous key stream. The paper showed that this was not the case. Some bits of the key had a bigger effect than others. Some bits had no effect at all (on the first few bytes of key stream). This is bad for two reasons. First, if you reduce the number of effective bits, it is easier to attack the keys. Second, the first few bytes of plaintext are usually easier to guess. For example, in WEP it is usually the LLC header that starts with the same hexadecimal value "AA". If you know the plaintext, you can derive the key stream and start attacking the key.

There is a very simple way to avoid the weakness: Discard the first few bytes of the RC4 key stream. In other words, wait until the RC4 algorithm gets going before starting to use the output. A recommendation from RSA Labs is to discard the first 256 bytes of the key stream, but of course WEP does not do this and such a change would mean that old systems would no longer interoperate.

You might think that this is not so bad. After all, you might not be using a weak key; or if you know which keys are weak, you could avoid them, right? Think again. Remember the IV is added to the secret key. And the IV is always changing. So sooner or later, the IV guarantees that a weak key is generated. It brings tears to your eyes, doesn't it! But there's worse to come.

Direct Key Attacks

In their landmark paper, Fluhrer et al. showed that using a public IV value appended to the secret key generated a huge weakness because it allowed the attacker to wait for a potentially weak key and directly attack the key. There are two cases, one in which the IV is appended (after the secret key) and one in which the IV is prepended (before the secret key). The prepend case is the more vulnerable?and it's the relevant case for WEP, which, by now, should come as no surprise to you.

If you are interested in how the attacks work, get a college degree in mathematics and read the paper. But in overview, the idea is based on exploiting the weak key problem in the first bytes. First assume that you know the plaintext for the first few bytes, which you do for IEEE 802.11 because it is usually an IEEE 802.1LLC SNAP header. Watch transmissions looking for a weak key generated by the IV. Now you know that there is a correlation between the ciphertext, the plaintext, and the secret key bytes. There are only a limited number of possible values for the first secret key byte that could match the plaintext and ciphertext. After capturing about 60 such messages, the attacker can guess the first key byte with reasonable certainty.

The method can be tuned to attack each secret key byte in turn so eventually the entire secret key can be extracted. Note that increasing the key size from 40 bits to 104 bits only means that it takes 2.5 times longer to extract the key?in other words, the time to crack the key goes up linearly with key size rather than exponentially.

All the previous weaknesses of WEP pale into insignificance compared to this attack. Remember that extracting the keys is the ultimate goal of an attacker, and here is a method that directly extracts the keys in linear time. This attack blew apart the remnants of WEP security. Worse, because it used a fairly mechanical approach, it was feasible to create a script tool that would do the job unattended.

Within months, some "helpful" person invested their time into generating a cracker tool. Publicizing the threat was a service to everyone, but I leave it as an exercise for the readers to determine what satisfaction is obtained by the authors of tools that turn threat into reality and lay waste to millions of dollars of investment. However, the tool was published, it is available on the Internet, and attackers can use it to crack WEP systems open at will.

    Part II: The Design of Wi-Fi Security