This paper describes an Internet security attack that could endanger the privacy of World Wide Web users and the integrity of their data. The attack can be carried out on today’s systems, endangering users of the most common Web browsers, including Netscape Navigator and Microsoft Internet Explorer.

Web spoofing allows an attacker to create a "shadow copy" of the entire World Wide Web. Accesses to the shadow Web are funneled through the attacker’s machine, allowing the attacker to monitor the all of the victim’s activities including any passwords or account numbers the victim enters. The attacker can also cause false or misleading data to be sent to Web servers in the victim’s name, or to the victim in the name of any Web server. In short, the attacker observes and controls everything the victim does on the Web.

We have implemented a demonstration version of this attack.

In a spoofing attack, the attacker creates misleading context in order to trick the victim into making an inappropriate security-relevant decision. A spoofing attack is like a con game: the attacker sets up a false but convincing world around the victim. The victim does something that would be appropriate if the false world were real. Unfortunately, activities that seem reasonable in the false world may have disastrous effects in the real world.

Spoofing attacks are possible in the physical world as well as the electronic one. For example, there have been several incidents in which criminals set up bogus automated-teller machines, typically in the public areas of shopping malls. The machines would accept ATM cards and ask the person to enter their PIN code. Once the machine had the victim’s PIN, it could either eat the card or "malfunction" and return the card. In either case, the criminals had enough information to copy the victim’s card and use the duplicate. In these attacks, people were fooled by the context they saw: the location of the machines, their size and weight, the way they were decorated, and the appearance of their electronic displays.

People using computer systems often make security-relevant decisions based on contextual cues they see. For example, you might decide to type in your bank account number because you believe you are visiting your bank’s Web page. This belief might arise because the page has a familiar look, because the bank’s URL appears in the browser’s location line, or for some other reason.

To appreciate the range and severity of possible spoofing attacks, we must look more deeply into two parts of the definition of spoofing: security-relevant decisions and context.

By "security-relevant decision," we mean any decision a person makes that might lead to undesirable results such as a breach of privacy or unauthorized tampering with data. Deciding to divulge sensitive information, for example by typing in a password or account number, is one example of a security-relevant decision. Choosing to accept a downloaded document is a security-relevant decision, since in many cases a downloaded document is capable of containing malicious elements that harm the person receiving the document.

Even the decision to accept the accuracy of information displayed by your computer can be security-relevant. For example, if you decide to buy a stock based on information you get from an online stock ticker, you are trusting that the information provided by the ticker is correct. If somebody could present you with incorrect stock prices, they might cause you to engage in a transaction that you would not have otherwise made, and this could cost you money.

A browser presents many types of context that users might rely on to make decisions. The text and pictures on a Web page might give some impression about where the page came from; for example, the presence of a corporate logo implies that the page originated at a certain corporation.

The appearance of an object might convey a certain impression; for example, neon green text on a purple background probably came from Wired magazine. You might think you’re dealing with a popup window when what you are seeing is really just a rectangle with a border and a color different from the surrounding parts of the screen. Particular graphical items like file-open dialog boxes are immediately recognized as having a certain purpose. Experienced Web users react to such cues in the same way that experienced drivers react to stop signs without reading them.

The names of objects can convey context. People often deduce what is in a file by its name. Is manual.doc the text of a user manual? (It might be another kind of document, or it might not be a document at all.) URLs are another example. Is MICR0S0FT.COM the address of a large software company? (For a while that address pointed to someone else entirely. By the way, the round symbols in MICR0S0FT here are the number zero, not the letter O.) Was dole96.org Bob Dole’s 1996 presidential campaign? (It was not; it pointed to a parody site.)

People often get context from the timing of events. If two things happen at the same time, you naturally think they are related. If you click over to your bank’s page and a username/password dialog box appears, you naturally assume that you should type the name and password that you use for the bank. If you click on a link and a document immediately starts downloading, you assume that the document came from the site whose link you clicked on. Either assumption could be wrong.

If you only see one browser window when an event occurs, you might not realize that the event was caused by another window hiding behind the visible one.

Modern user-interface designers spend their time trying to devise contextual cues that will guide people to behave appropriately, even if they do not explicitly notice the cues. While this is usually beneficial, it can become dangerous when people are accustomed to relying on context that is not always correct.

Another class of spoofing attack, which we will not discuss here, tricks the user’s software into an inappropriate action by presenting misleading information to that software. Examples of such attacks include TCP spoofing, in which Internet packets are sent with forged return addresses, and DNS spoofing, in which the attacker forges information about which machine names correspond to which network addresses. These other spoofing attacks are well known, so we will not discuss them further.

Web spoofing is a kind of electronic con game in which the attacker creates a convincing but false copy of the entire World Wide Web. The false Web looks just like the real one: it has all the same pages and links. However, the attacker controls the false Web, so that all network traffic between the victim’s browser and the Web goes through the attacker.

Since the attacker can observe or modify any data going from the victim to Web servers, as well as controlling all return traffic from Web servers to the victim, the attacker has many possibilities. These include surveillance and tampering.

Surveillance The attacker can passively watch the traffic, recording which pages the victim visits and the contents of those pages. When the victim fills out a form, the entered data is transmitted to a Web server, so the attacker can record that too, along with the response sent back by the server. Since most on-line commerce is done via forms, this means the attacker can observe any account numbers or passwords the victim enters.

As we will see below, the attacker can carry out surveillance even if the victim has a "secure" connection (usually via Secure Sockets Layer) to the server, that is, even if the victim’s browser shows the secure-connection icon (usually an image of a lock or a key).

Tampering The attacker is also free to modify any of the data traveling in either direction between the victim and the Web. The attacker can modify form data submitted by the victim. For example, if the victim is ordering a product on-line, the attacker can change the product number, the quantity, or the ship-to address.

The attacker can also modify the data returned by a Web server, for example by inserting misleading or offensive material in order to trick the victim or to cause antagonism between the victim and the server.

You may think it is difficult for the attacker to spoof the entire World Wide Web, but it is not. The attacker need not store the entire contents of the Web. The whole Web is available on-line; the attacker’s server can just fetch a page from the real Web when it needs to provide a copy of the page on the false Web.

The key to this attack is for the attacker’s Web server to sit between the victim and the rest of the Web. This kind of arrangement is called a "man in the middle attack" in the security literature.

The attacker’s first trick is to rewrite all of the URLs on some Web page so that they point to the attacker’s server rather than to some real server. Assuming the attacker’s server is on the machine www.attacker.org, the attacker rewrites a URL by adding http://www.attacker.org to the front of the URL. For example, http://home.netscape.com becomes http://www.attacker.org/http://home.netscape.com. (The URL rewriting technique has been used for other reasons by two other Web sites, the Anonymizer and the Zippy filter. See page 9 for details.)

Figure 1 shows what happens when the victim requests a page through one of the rewritten URLs. The victim’s browser requests the page from www.attacker.org, since the URL starts with http://www.attacker.org. The remainder of the URL tells the attacker’s server where on the Web to go to get the real document.

Once the attacker’s server has fetched the real document needed to satisfy the request, the attacker rewrites all of the URLs in the document into the same special form by splicing http://www.attacker.org/ onto the front. Then the attacker’s server provides the rewritten page to the victim’s browser.

Since all of the URLs in the rewritten page now point to www.attacker.org, if the victim follows a link on the new page, the page will again be fetched through the attacker’s server. The victim remains trapped in the attacker’s false Web, and can follow links forever without leaving it.

If the victim fills out a form on a page in a false Web, the result appears to be handled properly. Spoofing of forms works naturally because forms are integrated closely into the basic Web protocols: form submissions are encoded in URLs and the replies are ordinary HTML Since any URL can be spoofed, forms can also be spoofed.

When the victim submits a form, the submitted data goes to the attacker’s server. The attacker’s server can observe and even modify the submitted data, doing whatever malicious editing desired, before passing it on to the real server. The attacker’s server can also modify the data returned in response to the form submission.

One distressing property of this attack is that it works even when the victim requests a page via a "secure" connection. If the victim does a "secure" Web access ( a Web access using the Secure Sockets Layer) in a false Web, everything will appear normal: the page will be delivered, and the secure connection indicator (usually an image of a lock or key) will be turned on.

The victim’s browser says it has a secure connection because it does have one. Unfortunately the secure connection is to www.attacker.org and not to the place the victim thinks it is. The victim’s browser thinks everything is fine: it was told to access a URL at www.attacker.org so it made a secure connection to www.attacker.org. The secure-connection indicator only gives the victim a false sense of security.

To start an attack, the attacker must somehow lure the victim into the attacker’s false Web. There are several ways to do this. An attacker could put a link to a false Web onto a popular Web page. If the victim is using Web-enabled email, the attacker could email the victim a pointer to a false Web, or even the contents of a page in a false Web. Finally, the attacker could trick a Web search engine into indexing part of a false Web.

The attack as described thus far is fairly effective, but it is not perfect. There is still some remaining context that can give the victim clues that the attack is going on. However, it is possible for the attacker to eliminate virtually all of the remaining clues of the attack’s existence.

Such evidence is not too hard to eliminate because browsers are very customizable. The ability of a Web page to control browser behavior is often desirable, but when the page is hostile it can be dangerous.

The status line is a single line of text at the bottom of the browser window that displays various messages, typically about the status of pending Web transfers.

The attack as described so far leaves two kinds of evidence on the status line. First, when the mouse is held over a Web link, the status line displays the URL the link points to. Thus, the victim might notice that a URL has been rewritten. Second, when a page is being fetched, the status line briefly displays the name of the server being contacted. Thus, the victim might notice that www.attacker.org is displayed when some other name was expected.

The attacker can cover up both of these cues by adding a JavaScript program to every rewritten page. Since JavaScript programs can write to the status line, and since it is possible to bind JavaScript actions to the relevant events, the attacker can arrange things so that the status line participates in the con game, always showing the victim what would have been on the status line in the real Web. Thus the spoofed context becomes even more convincing.

The browser’s location line displays the URL of the page currently being shown. The victim can also type a URL into the location line, sending the browser to that URL. The attack as described so far causes a rewritten URL to appear in the location line, giving the victim a possible indication that an attack is in progress.

This clue can be hidden using JavaScript. A JavaScript program can hide the real location line and replace it by a fake location line which looks right and is in the expected place. The fake location line can show the URL the victim expects to see. The fake location line can also accept keyboard input, allowing the victim to type in URLs normally. Typed-in URLs can be rewritten by the JavaScript program before being accessed.

There is one clue that the attacker cannot eliminate, but it is very unlikely to be noticed.

By using the browser’s "view source" feature, the victim can look at the HTML source for the currently displayed page. By looking for rewritten URLs in the HTML source, the victim can spot the attack. Unfortunately, HTML source is hard for novice users to read, and very few Web surfers bother to look at the HTML source for documents they are visiting, so this provides very little protection.

A related clue is available if the victim chooses the browser’s "view document information" menu item. This will display information including the document’s real URL, possibly allowing the victim to notice the attack. As above, this option is almost never used so it is very unlikely that it will provide much protection.

There are several ways the victim might accidentally leave the attacker’s false Web during the attack. Accessing a bookmark or jumping to a URL by using the browser’s "Open location" menu item might lead the victim back into the real Web. The victim might then reenter the false Web by clicking the "Back" button. We can imagine that the victim might wander in and out of one or more false Webs. Of course, bookmarks can also work against the victim, since it is possible to bookmark a page in a false Web. Jumping to such a bookmark would lead the victim into a false Web again.

Some people have suggested that this attack can be deterred by finding and punishing the attacker. It is true that the attacker’s server must reveal its location in order to carry out the attack, and that evidence of that location will almost certainly be available after an attack is detected.

Unfortunately, this will not help much in practice because attackers will break into the machine of some innocent person and launch the attack there. Stolen machines will be used in these attacks for the same reason most bank robbers make their getaways in stolen cars.

The URL-rewriting part of our demonstration program is based on Henry Minsky’s code for the Zippy filter. We are grateful to David Hopwood for useful discussions about spoofing attacks, and to Gary McGraw and Laura Felten for comments on drafts of this paper. The figure was designed by Gary McGraw.

More information is available from our Web page at http://www.cs.princeton.edu/sip, or from Prof. Edward Felten at felten@cs.princeton.edu or (609) 258-5906.