Didier Stevens

This starts a series of post leading up to my PDF talk at the next Belgian ISSA and OWASP chapter event. I’ll be publishing a couple of my PDF tools.

Next video shows how I use my PDF parser to analyze a malicious PDF file, and extract the shell code.

Searching for keyword javascript yields 2 indirect objects referencing /JavaScript objects. The JavaScript is executed through an automatic annotation (/AA) when the page is rendered (e.g. when the PDF document is opened, as it contains only one page). Decompressing the second /JavaScript object (34) displays the code.

collectEmailInfo is an undocument Adobe Acrobat JavaScript method with a vulnerability (fixed in Adobe Acrobat Reader 8.1.2). My Spidermonkey helps me to extract the shell code.

YouTube, Vimeo and hires Xvid.

A PDF stream object is a sequence of bytes. There is a virtually unlimited number of ways to represent the same byte sequence. After Names and Strings obfuscation, let’s take a look at streams.

A PDF stream object is composed of a dictionary (<< >>), the keyword stream, a sequence of bytes and the keyword endstream. All streams must be indirect objects. Here is an example:

This stream is indirect object 5 version 0. The stream dictionary must have a /Length entry, to document the length of the (encoded) byte sequence. The stream and endstream keywords are terminated with the EOL character(s). In this example, the byte sequence is a set of instructions for the PDF reader to render the string Hello World with a given font at a precise position. It’s precisely 42 bytes long.

In this example, the byte sequence is represented literally, but it’s possible (and usual) to encode the byte sequence. This is done with a stream filter. A stream filter specifies how the sequence of bytes has to be decoded. Let’s take the same example, but with an ASCII85 encoding:

The /Filter entry instructs the PDF reader how to decode the byte sequence (/ASCII85Decode). Notice the change of the length value. There are many encoding schemes (ASCII filters and decompression filters), here is a list:

• ASCIIHexDecode
• ASCII85Decode
• LZWDecode
• FlateDecode
• RunLengthDecode
• CCITTFaxDecode
• JBIG2Decode
• DCTDecode
• JPXDecode
• Crypt

This list is not so long, so why do I claim an almost limitless number of ways to encode a stream? I have 2 reasons:

1. Many filters, like /FlateDecode, take parameters (in this case, the compression level), which influence the encoding too
2. Filters can be cascaded, meaning that the stream has to be decoded by more than one filter

Here is our example, where the stream is encoded twice, first with ASCII85 and then with plain HEX (I know, this is rather pointless, but it yields simple and readable examples):

Cascading filters also inspired me to create a couple of test PDF documents. For example, I’ve created a 2642 bytes small PDF document that contains a 1GB large stream (a ZIP bomb of sorts). Some PDF readers will choke on this document.

In this post, I show how basic features of the PDF language can be used to generate polymorphic variants of (malicious) PDF documents. If you code a PDF parser, write signatures (AV, IDS, …) or analyze (malicious) PDF documents, you should to be aware of these features.

Official language specifications are interesting documents, I used to read them from front to back. I especially appreciate the inclusion of a formal language description, for example in Backus–Naur form. But nowadays, I don’t take the time to do this anymore.

While browsing through the official PDF documentation, I took particular interest in the rules to express lexemes. There are many ways to write the same token, offering opportunities to evade known-pattern recognition systems, like AV and NIDS.

Building a test file

Before I show some examples, let’s build a test PDF file that will start the default browser and navigate to a site each time the document is opened.

Opening a web page from a PDF file can be done with an URI action, like this:

This is the same type of object used in the malicious mailto PDF files.

An action must be triggered by an event, examples of such triggers are the association of an action to the display of a page or the opening of the PDF document. We will use the OpenAction to trigger our URI action object each time our test PDF document is opened:

I add the URI action object and the OpenAction event to the hello world PDF file I used in a previous post, to build our test PDF. You can download all examples here. Opening the test PDF document launches IE:

Now that we have our test PDF, let’s look at the ways we can change its representation without changing its rendering. This is what I’m covering (this list is not exhaustive):

• Names
  • Hexadecimal encoding
• Strings
  • Newline escaping
  • Octal encoding
  • Hexadecimal encoding
  • Hexadecimal whitespacing
  • Encryption

Name representation

The tokens preceded by a / (slash) in the URI action object are called Names in the official PDF description. Names are case-sensitive. The characters used in a Name are limited to a specific set, but since PDF specification version 1.2, a lexical convention has been added to represent a character with its hexadecimal ANSI-code, like this #XX.

This allows use to rewrite the /URI name in several ways, for example: #55RI.

Or #55#52#49

Pattern matching algorithms must take into account these different representations to successfully match a pattern. A standard way to deal with this is canonicalization. First, the token is reduced to a canonical form (e.g. replace all #xx representations by the character they stand for), and second, pattern matching is performed on the canonical form.

String representation

Strings too can be represented in many forms. One way to represent strings, is to type the text between parentheses:

Splitting strings over several lines can be done by adding a backslash (\) at the end of each line:

Of course, we are not limited by the numbers of lines, we can add a backslash after each character:

A character in a string can be represented by its octal code, like this:

And this can be done for every character in the string:

One more way to represent a string, is hexadecimal:

You’re allowed to put whitespace between the hex digits:

And you’re not limited in the amount of whitespace you use:

This whitespace usage reminds me of the IE zero-byte trick in html.

I want to finish this long list of examples with PDF encryption. One more way to change the representation of a PDF document is encryption. PDFs can be encrypted without requiring the user to provide a password to view the encrypted document, this form of encryption is used for DRM. Ever had a PDF with printing or text copy disabled? That’s an encrypted PDF.

When a PDF is encrypted, only the strings and streams are encrypted, the objects themselves are not encrypted. Encrypted strings are one more way to change the representation of a string.

Here’s an example:

I know that PDF encryption has already been used to mislead SPAM filters.

Final thoughts

These many features of the PDF language providing flexibility in representation of names and strings, can also be used to generate polymorphic forms of the same malicious PDF. If you need to scan PDF documents, you need to be aware of all these features and have tools that support them.

There are indications that most AV products don’t canonicalize PDF documents prior to signature matching. I did some tests with a malicious mailto PDF document, and changing the string representation of the mailto URI action using the hexadecimal forms allows AV detection evasion. Adding whitespace wasn’t necessary, switching to hex was enough. The ClamAV source code for PDF documents has more evidence of PDF canonicalization issues in AV software, here is a string compare for the Length name without canicalization:

This will not match if hex codes are used (#).

I tested all my examples with Adobe Acrobat Reader 8.1.2 and Foxit Reader 2.2 without problems. But Foxit Reader 2.2 gave me an unpleasant surprise, more on this in a next post.

I wonder if malicious PDF samples will be used in the Race to Zero.