3. The Internet Threat Model

A THREAT MODEL describes the capabilities that an attacker is assumed to be able to deploy against a resource. It should contain such information as the resources available to an attacker in terms of information, computing capability, and control of the system. The purpose of a threat model is twofold. First, we wish to identify the threats we are concerned with. Second, we wish to rule some threats explicitly out of scope. Nearly every security system is vulnerable to a sufficiently dedicated and resourceful attacker.

The Internet environment has a fairly well understood threat model. In general, we assume that the end-systems engaging in a protocol exchange have not themselves been compromised. Protecting against an attack when one of the end-systems has been compromised is extraordinarily difficult. It is, however, possible to design protocols which minimize the extent of the damage done under these circumstances.

By contrast, we assume that the attacker has nearly complete control of the communications channel over which the end-systems communicate. This means that the attacker can read any PDU (Protocol Data Unit) on the network and undetectably remove, change, or inject forged packets onto the wire. This includes being able to generate packets that appear to be from a trusted machine. Thus, even if the end-system with which you wish to communicate is itself secure, the Internet environment provides no assurance that packets which claim to be from that system in fact are.

It's important to realize that the meaning of a PDU is different at different levels. At the IP level, a PDU means an IP packet. At the TCP level, it means a TCP segment. At the application layer, it means some kind of application PDU. For instance, at the level of email, it might either mean an RFC-822 message or a single SMTP command. At the HTTP level, it might mean a request or response.

3.1. Limited Threat Models

As we've said, a resourceful and dedicated attacker can control the entire communications channel. However, a large number of attacks can be mounted by an attacker with fewer resources. A number of currently known attacks can be mounted by an attacker with limited control of the network. For instance, password sniffing attacks can be mounted by an attacker who can only read arbitrary packets. This is generally referred to as a PASSIVE ATTACK [INTAUTH].

By contrast, Morris' sequence number guessing attack [SEQNUM] can be mounted by an attacker who can write but not read arbitrary packets. Any attack which requires the attacker to write to the network is known as an ACTIVE ATTACK.

Thus, a useful way of organizing attacks is to divide them based on the capabilities required to mount the attack. The rest of this section describes these categories and provides some examples of each category.

3.2. Passive Attacks

In a passive attack, the attacker reads packets off the network but does not write them. The simplest way to mount such an attack is to simply be on the same LAN as the victim. On most common LAN configurations, including Ethernet, 802.3, and FDDI, any machine on the wire can read all traffic destined for any other machine on the same LAN. Note that switching hubs make this sort of sniffing substantially more difficult, since traffic destined for a machine only goes to the network segment which that machine is on.

Similarly, an attacker who has control of a host in the communications path between two victim machines is able to mount a passive attack on their communications. It is also possible to compromise the routing infrastructure to specifically arrange that traffic passes through a compromised machine. This might involve an active attack on the routing infrastructure to facilitate a passive attack on a victim machine.

Wireless communications channels deserve special consideration, especially with the recent and growing popularity of wireless-based LANs, such as those using 802.11. Since the data is simply broadcast on well known radio frequencies, an attacker simply needs to be able to receive those transmissions. Such channels are especially vulnerable to passive attacks. Although many such channels include cryptographic protection, it is often of such poor quality as to be nearly useless [WEP].

In general, the goal of a passive attack is to obtain information which the sender and receiver would prefer to remain private. This private information may include credentials useful in the electronic world and/or passwords or credentials useful in the outside world, such as confidential business information.

3.2.1. Confidentiality Violations

The classic example of passive attack is sniffing some inherently private data off of the wire. For instance, despite the wide availability of SSL, many credit card transactions still traverse the Internet in the clear. An attacker could sniff such a message and recover the credit card number, which can then be used to make fraudulent transactions. Moreover, confidential business information is routinely transmitted over the network in the clear in email.

3.2.2. Password Sniffing

Another example of a passive attack is PASSWORD SNIFFING. Password sniffing is directed towards obtaining unauthorized use of resources. Many protocols, including [TELNET], [POP], and [NNTP] use a shared password to authenticate the client to the server. Frequently, this password is transmitted from the client to the server in the clear over the communications channel. An attacker who can read this traffic can therefore capture the password and REPLAY it. In other words, the attacker can initiate a connection to the server and pose as the client and login using the captured password.

Note that although the login phase of the attack is active, the actual password capture phase is passive. Moreover, unless the server checks the originating address of connections, the login phase does not require any special control of the network.

3.2.3. Offline Cryptographic Attacks

Many cryptographic protocols are subject to OFFLINE ATTACKS. In such a protocol, the attacker recovers data which has been processed using the victim's secret key and then mounts a cryptanalytic attack on that key. Passwords make a particularly vulnerable target because they are typically low entropy. A number of popular password-based challenge response protocols are vulnerable to DICTIONARY ATTACK. The attacker captures a challenge-response pair and then proceeds to try entries from a list of common words (such as a dictionary file) until he finds a password that produces the right response.

A similar such attack can be mounted on a local network when NIS is used. The Unix password is crypted using a one-way function, but tools exist to break such crypted passwords [KLEIN]. When NIS is used, the crypted password is transmitted over the local network and an attacker can thus sniff the password and attack it.

Historically, it has also been possible to exploit small operating system security holes to recover the password file using an active attack. These holes can then be bootstrapped into an actual account by using the aforementioned offline password recovery techniques. Thus we combine a low-level active attack with an offline passive attack.

3.3. Active Attacks

When an attack involves writing data to the network, we refer to this as an ACTIVE ATTACK. When IP is used without IPsec, there is no authentication for the sender address. As a consequence, it's straightforward for an attacker to create a packet with a source address of his choosing. We'll refer to this as a SPOOFING ATTACK.

Under certain circumstances, such a packet may be screened out by the network. For instance, many packet filtering firewalls screen out all packets with source addresses on the INTERNAL network that arrive on the EXTERNAL interface. Note, however, that this provides no protection against an attacker who is inside the firewall. In general, designers should assume that attackers can forge packets.

However, the ability to forge packets does not go hand in hand with the ability to receive arbitrary packets. In fact, there are active attacks that involve being able to send forged packets but not receive the responses. We'll refer to these as BLIND ATTACKS.

Note that not all active attacks require forging addresses. For instance, the TCP SYN denial of service attack [TCPSYN] can be mounted successfully without disguising the sender's address. However, it is common practice to disguise one's address in order to conceal one's identity if an attack is discovered.

Each protocol is susceptible to specific active attacks, but experience shows that a number of common patterns of attack can be adapted to any given protocol. The next sections describe a number of these patterns and give specific examples of them as applied to known protocols.

3.3.1. Replay Attacks

In a REPLAY ATTACK, the attacker records a sequence of messages off of the wire and plays them back to the party which originally received them. Note that the attacker does not need to be able to understand the messages. He merely needs to capture and retransmit them.

For example, consider the case where an S/MIME message is being used to request some service, such as a credit card purchase or a stock trade. An attacker might wish to have the service executed twice, if only to inconvenience the victim. He could capture the message and replay it, even though he can't read it, causing the transaction to be executed twice.

3.3.2. Message Insertion

In a MESSAGE INSERTION attack, the attacker forges a message with some chosen set of properties and injects it into the network. Often this message will have a forged source address in order to disguise the identity of the attacker.

For example, a denial-of-service attack can be mounted by inserting a series of spurious TCP SYN packets directed towards the target host. The target host responds with its own SYN and allocates kernel data structures for the new connection. The attacker never completes the 3-way handshake, so the allocated connection endpoints just sit there taking up kernel memory. Typical TCP stack implementations only

allow some limited number of connections in this "half-open" state and when this limit is reached, no more connections can be initiated, even from legitimate hosts. Note that this attack is a blind attack, since the attacker does not need to process the victim's SYNs.

3.3.3. Message Deletion

In a MESSAGE DELETION attack, the attacker removes a message from the wire. Morris' sequence number guessing attack [SEQNUM] often requires a message deletion attack to be performed successfully. In this blind attack, the host whose address is being forged will receive a spurious TCP SYN packet from the host being attacked. Receipt of this SYN packet generates a RST, which would tear the illegitimate connection down. In order to prevent this host from sending a RST so that the attack can be carried out successfully, Morris describes flooding this host to create queue overflows such that the SYN packet is lost and thus never responded to.

3.3.4. Message Modification

In a MESSAGE MODIFICATION attack, the attacker removes a message from the wire, modifies it, and reinjects it into the network. This sort of attack is particularly useful if the attacker wants to send some of the data in the message but also wants to change some of it.

Consider the case where the attacker wants to attack an order for goods placed over the Internet. He doesn't have the victim's credit card number so he waits for the victim to place the order and then replaces the delivery address (and possibly the goods description) with his own. Note that this particular attack is known as a CUT- AND-PASTE attack since the attacker cuts the credit card number out of the original message and pastes it into the new message.

Another interesting example of a cut-and-paste attack is provided by [IPSPPROB]. If IPsec ESP is used without any MAC then it is possible for the attacker to read traffic encrypted for a victim on the same machine. The attacker attaches an IP header corresponding to a port he controls onto the encrypted IP packet. When the packet is received by the host it will automatically be decrypted and forwarded to the attacker's port. Similar techniques can be used to mount a session hijacking attack. Both of these attacks can be avoided by always using message authentication when you use encryption. Note that this attack only works if (1) no MAC check is being used, since this attack generates damaged packets (2) a host-to-host SA is being used, since a user-to-user SA will result in an inconsistency between the port associated with the SA and the target port. If the receiving machine is single-user than this attack is infeasible.

3.3.5. Man-In-The-Middle

A MAN-IN-THE-MIDDLE attack combines the above techniques in a special form: The attacker subverts the communication stream in order to pose as the sender to receiver and the receiver to the sender:

What Alice and Bob think: Alice <----------------------------------------------> Bob

What's happening: Alice <----------------> Attacker <----------------> Bob

This differs fundamentally from the above forms of attack because it attacks the identity of the communicating parties, rather than the data stream itself. Consequently, many techniques which provide integrity of the communications stream are insufficient to protect against man-in-the-middle attacks.

Man-in-the-middle attacks are possible whenever a protocol lacks PEER ENTITY AUTHENTICATION. For instance, if an attacker can hijack the client TCP connection during the TCP handshake (perhaps by responding to the client's SYN before the server does), then the attacker can open another connection to the server and begin a man-in-the-middle attack. It is also trivial to mount man-in-the-middle attacks on local networks via ARP spoofing -- the attacker forges an ARP with the victim's IP address and his own MAC address. Tools to mount this sort of attack are readily available.

Note that it is only necessary to authenticate one side of the transaction in order to prevent man-in-the-middle attacks. In such a situation the the peers can establish an association in which only one peer is authenticated. In such a system, an attacker can initiate an association posing as the unauthenticated peer but cannot transmit or access data being sent on a legitimate connection. This is an acceptable situation in contexts such as Web e-commerce where only the server needs to be authenticated (or the client is independently authenticated via some non-cryptographic mechanism such as a credit card number).

3.4. Topological Issues

In practice, the assumption that it's equally easy for an attacker to read and generate all packets is false, since the Internet is not fully connected. This has two primary implications.

3.5. On-path versus off-path

In order for a datagram to be transmitted from one host to another, it generally must traverse some set of intermediate links and gateways. Such gateways are naturally able to read, modify, or remove any datagram transmitted along that path. This makes it much easier to mount a wide variety of attacks if you are on-path.

Off-path hosts can, of course, transmit arbitrary datagrams that appear to come from any hosts but cannot necessarily receive datagrams intended for other hosts. Thus, if an attack depends on being able to receive data, off-path hosts must first subvert the topology in order to place themselves on-path. This is by no means impossible but is not necessarily trivial.

Applications protocol designers MUST NOT assume that all attackers will be off-path. Where possible, protocols SHOULD be designed to resist attacks from attackers who have complete control of the network. However, designers are expected to give more weight to attacks which can be mounted by off-path attackers as well as on-path ones.

3.6. Link-local

One specialized case of on-path is being on the same link. In some situations, it's desirable to distinguish between hosts who are on the local network and those who are not. The standard technique for this is verifying the IP TTL value [IP]. Since the TTL must be decremented by each forwarder, a protocol can demand that TTL be set to 255 and that all receivers verify the TTL. A receiver then has some reason to believe that conforming packets are from the same link. Note that this technique must be used with care in the presence of tunneling systems, since such systems may pass packets without decrementing TTL.