14. Unicode issues

14.1. Security vulnerabilities

14.1.1. Special characters

Fullwidth (U+FF01—U+FF60) and halfwidth (U+FF61—U+FFEE) characters have been used in 2007 to bypass security checks. Examples with the Unicode normalization:

  • U+FF0E is normalized to . (U+002E) in NFKC
  • U+FF0F is normalized to / (U+002F) in NFKC

Some important characters have also “alternatives” in Unicode:

  • Windows directory separator, \ (U+005C): U+20E5, U+FF3C
  • UNIX directory separator, / (U+002F): U+2215, U+FF0F
  • Parent directory, .. (U+002E, U+002E): U+FF0E

For more information, read GS07-01 Full-Width and Half-Width Unicode Encoding IDS/IPS/WAF Bypass Vulnerability (GamaTEAM, april 2007).

14.1.2. Non-strict UTF-8 decoder: overlong byte sequences and surrogates

An UTF-8 decoder has to reject overlong byte sequences, or an attacker can use them to bypass security checks (e.g. check rejecting string containing nul bytes, 0x00). For example, 0xC0 0x80 byte sequence must raise an error and not be decoded as U+0000, and ”.” (U+002E) can be encoded to 0xC0 0xAE (two bytes instead of one) to bypass directory traversal checks.

Surrogates characters are also invalid in UTF-8: characters in U+D800—U+DFFF have to be rejected. See the table 3-7 in the Conformance chapter of the Unicode standard (december 2009); and the section 3 (UTF-8 definition) of UTF-8, a transformation format of ISO 10646 (RFC 3629, november 2003).

The libxml2 library had such vulnerability until january 2008: CVE-2007-6284.

Some PHP functions use a strict UTF-8 decoder (e.g. mb_convert_encoding()), some other don’t. For example, utf8_decode() and mb_strlen() accept 0xC0 0x80 in PHP 5.3.2. The UTF-8 decoder of Python 3 is strict, whereas the UTF-8 decoder of Python 2 accepts surrogates (to keep the backward compatibility). In Python 3, the error handler surrogatepass can be used to encode and decode surrogates.


The Java and Tcl languages use a variant of UTF-8 which encodes the nul character (U+0000) as the overlong byte sequence 0xC0 0x80, instead of 0x00, for practical reasons.

14.1.3. Check byte strings before decoding them to character strings

Some applications check user inputs as byte strings, but then process them as character strings. This vulnerability can be used to bypass security checks.

The WordPress blog tool had such issue with PHP5 and MySQL: WordPress Charset SQL Injection Vulnerability (Abel Cheung, december 2007). WordPress used the PHP function addslashes() on the input byte strings. This function adds 0x5C prefix to 0x00, 0x22, 0x27 and 0x5C bytes. If a input string is encoded to ISO 8859-1, this operation escapes a quote: ' (U+0027) becomes \' ({U+005C, U+0027}).

The problem is that addslashes() process byte strings, whereas the result is used by MySQL which process character strings. Example with Big5 encoding: 0xB5 0x27 cannot be decoded from Big5, but escaped it becomes 0xB5 0x5C 0x27 which is decoded to {U+8A31, U+0027}. The 0x5C byte is no more a backslash: it is part of the multibyte character U+8A31 encoded to 0xB5 0x5C. The solution is to use mysql_real_escape_string() function, instead of addslashes(), which process inputs as character strings using the MySQL connection encoding.

See also

CVE-2006-2314 (PostgreSQL, may 2006), CVE-2006-2753 (MySQL, may 2006) and CVE-2008-2384 (libapache2-mod-auth-mysql, january 2009).