Submitted By Breno Silva Pinto and Ryan Barnett
WAF Normalization and I18N
Web application firewalls must be able to handle Internationaliztion (I18N) and thus properly handle various data encodings including Unicode and UTF-8 in order to prevent not only evasion issues but also to minimize false positives. In an earlier blog post, we highlighted ModSecurity's new support for Unicode mapping and decoding. This capability helps us to more accurately decode characters from different Unicode code points. While this certainly helps our accuracy, we still had the issue of UTF-8 encodings.. This is a challenge for any WAF as it must be able to handle UTF-8 encodings of characters for different languages such as Portuguese. So, if you are running ModSecurity to protect a non-English language website then this blog post is for you! We introduce a new transformation function called utf8toUnicode that helps to normalize data for inspection.
Incorrect UTF-8 Decoding
We have received some recent reports of false positive issues with the OWASP ModSecurity CRS that were due to existence of UTF-8 encoded characters. As an example, the Portuguese language has many "special" characters with different accent characters. For example:
1- á Á ã Ã â Â à À
2- é É ê Ê í Í
3- ó Ó õ Õ ô Ô
4- ú Ú ü Ü
5- ç Ç
When these characters are UTF-8 encoded, they use multiple bytes. As an example, the "ę" character is encoded as "%c4%99". If ModSecurity only applies the standard t:urlDecodeUni, it will decode each byte individually which results in an impedance mismatch. In this case, this incorrect decoding resulted in a false positive match against some SQL Injection rules in the OWASP ModSecurity CRS. While this is a bit of a pain, it is not as bad as a false negative bypass situation that may be caused by this type of incorrect decoding . Let's look an this type of SQL Injection evasion issue. What if we send the following request:
http://172.16.51.132/index.php?foo='úníón+séléct+data+fróm+námés
Let's see how ModSecurity will decode this data when checking for an example SQL Injection keyword:
Recipe: Invoking rule 21b67e38;
Rule 21b67e38: SecRule "ARGS:foo""@rx select""phase:2,log,auditlog,pass,id:1111,t:urlDecodeUni"
T (0) urlDecodeUni:"'\xc3\xban\xc3\xad\xc3\xb3n s\xc3\xa9l\xc3\xa9ctdata fr\xc3\xb3m n\xc3\xa1m\xc3\xa9s"
Transformation completed in 13 usec.
Executing operator "rx" withparam "select" against ARGS:foo.
Target value:"'\xc3\xban\xc3\xad\xc3\xb3n s\xc3\xa9l\xc3\xa9ct data fr\xc3\xb3mn\xc3\xa1m\xc3\xa9s"
Operator completed in 7 usec.
Rulereturned 0.
As you can see above the character sequence "úníón+séléct+data+fróm+námés" was handled by the engine as "\xc3\xban\xc3\xad\xc3\xb3ns\xc3\xa9l\xc3\xa9ct data fr\xc3\xb3m n\xc3\xa1m\xc3\xa9s". However the rulewas looking for the pattern "select". So, what happen if the application applies best-fit mapping conversions and removes the accents before sending the data tothe database ? This may allow the payload to bypass our signatures.
Utf8 to Unicode Mapping
In order to better handle this data, we should first map the UTF-8 encoded data to Unicode and then use the unicode point mapping capabilities mentioned at the beginning of the blog post. This configuration is achieved by first setting the SecUnicodeCodePage and SecUnicodeMapFile directives in your main ModSecurity configuration file:
SecUnicodeCodePage 20127SecUnicodeMapFile /etc/apache2/unicode.mapping
The SecUnicodeCodePage directive sets the proper Unicode code point used for your site. The example 20127 is the US-ASCII code point. The SecUnicodeMapFile points to the unicode.mapping file that comes with the ModSecurity source archive and includes all of the Unicode conversions. Here is an example of the 20127 (US-ASCII) mapping data:
20127 (US-ASCII)00a0:20 00a1:21 00a2:63 00a4:24 00a5:59 00a6:7c 00a9:43 00aa:6100ab:3c 00ad:2d 00ae:52 00b2:32 00b3:33 00b7:2e 00b8:2c 00b9:31 00ba:6f 00bb:3e00c0:41 00c1:41 00c2:41 00c3:41 00c4:41 00c5:41 00c6:41 00c7:43 00c8:45 00c9:4500ca:45 00cb:45 00cc:49 00cd:49 00ce:49 00cf:49 00d0:44 00d1:4e 00d2:4f 00d3:4f00d4:4f 00d5:4f 00d6:4f 00d8:4f 00d9:55 00da:55 00db:55 00dc:55 00dd:59 00e0:6100e1:61 00e2:61 00e3:61 00e4:6100e5:61 00e6:61 00e7:63 00e8:6500e9:65 00ea:65 00eb:65 00ec:69 00ed:69 00ee:69 00ef:69 00f1:6e 00f2:6f 00f3:6f00f4:6f 00f5:6f 00f6:6f 00f8:6f 00f9:75 00fa:75 00fb:75 00fc:75 00fd:79 00ff:790100:41 0101:61 0102:41 0103:61 0104:41 0105:61 0106:43 0107:63 0108:43 0109:63010a:43 010b:63 010c:43 010d:63 010e:44 010f:64 0110:44 0111:64 0112:45 0113:650114:45 0115:65 0116:45 0117:65 0118:45 0119:65 011a:45 011b:65 011c:47 011d:67011e:47 011f:67 0120:47 0121:67 0122:47 0123:67 0124:48 0125:68 0126:48 0127:680128:49 0129:69 012a:49 012b:69 012c:49 012d:69 012e:49 012f:69 0130:49 0131:690134:4a 0135:6a 0136:4b 0137:6b 0139:4c 013a:6c 013b:4c 013c:6c 013d:4c 013e:6c0141:4c 0142:6c 0143:4e 0144:6e 0145:4e 0146:6e 0147:4e 0148:6e 014c:4f 014d:6f014e:4f 014f:6f 0150:4f 0151:6f 0152:4f 0153:6f 0154:52 0155:72 0156:52 0157:720158:52 0159:72 015a:53 015b:73 015c:53 015d:73 015e:53 015f:73 0160:53 0161:730162:54 0163:74 0164:54 0165:74 0166:54 0167:74 0168:55 0169:75 016a:55 016b:75016c:55 016d:75 016e:55 016f:75 0170:55 0171:75 0172:55 0173:75 0174:57 0175:770176:59 0177:79 0178:59 0179:5a 017b:5a 017c:7a 017d:5a 017e:7a 0180:62 0189:440191:46 0192:66 0197:49 019a:6c 019f:4f 01a0:4f 01a1:6f 01ab:74 01ae:54 01af:5501b0:75 01b6:7a 01cd:41 01ce:61 01cf:49 01d0:69 01d1:4f 01d2:6f 01d3:55 01d4:7501d5:55 01d6:75 01d7:55 01d8:75 01d9:55 01da:75 01db:55 01dc:75 01de:41 01df:6101e4:47 01e5:67 01e6:47 01e7:67 01e8:4b 01e9:6b 01ea:4f 01eb:6f 01ec:4f 01ed:6f01f0:6a 0261:67 02b9:27 02ba:22 02bc:27 02c4:5e 02c6:5e 02c8:27 02cb:60 02cd:5f02dc:7e 0300:60 0302:5e 0303:7e 030e:22 0331:5f 0332:5f 2000:20 2001:20 2002:202003:20 2004:20 2005:20 2006:20 2010:2d 2011:2d 2013:2d 2014:2d 2018:27 2019:27201a:2c 201c:22 201d:22 201e:22 2022:2e 2026:2e 2032:27 2035:60 2039:3c 203a:3e2122:54 ff01:21 ff02:22 ff03:23 ff04:24 ff05:25 ff06:26 ff07:27 ff08:28 ff09:29ff0a:2a ff0b:2b ff0c:2c ff0d:2d ff0e:2e ff0f:2f ff10:30 ff11:31 ff12:32 ff13:33ff14:34 ff15:35 ff16:36 ff17:37 ff18:38 ff19:39 ff1a:3a ff1b:3b ff1c:3c ff1d:3dff1e:3e ff20:40 ff21:41 ff22:42 ff23:43 ff24:44 ff25:45 ff26:46 ff27:47 ff28:48ff29:49 ff2a:4a ff2b:4b ff2c:4c ff2d:4d ff2e:4e ff2f:4f ff30:50 ff31:51 ff32:52ff33:53 ff34:54 ff35:55 ff36:56 ff37:57 ff38:58 ff39:59 ff3a:5a ff3b:5b ff3c:5cff3d:5d ff3e:5e ff3f:5f ff40:60 ff41:61 ff42:62 ff43:63 ff44:64 ff45:65 ff46:66ff47:67 ff48:68 ff49:69 ff4a:6a ff4b:6b ff4c:6c ff4d:6d ff4e:6e ff4f:6f ff50:70ff51:71 ff52:72 ff53:73 ff54:74 ff55:75 ff56:76 ff57:77 ff58:78 ff59:79 ff5a:7aff5b:7b ff5c:7c ff5d:7d ff5e:7e
With these Unicode mapping directives in place, we now can use the new t:utf8toUnicode transformation function which was just added to the v2.7.0 code. Here is a new example rule:
SecRule "ARGS:foo" "@rx select" "phase:2,log,auditlog,pass,id:1111,t:utf8toUnicode,t:urlDecodeUni"
Let's see how ModSecurity will handle the same request, but nowusing the mentioned features:
Recipe: Invoking rule 21717d58;
Rule 21717d58: SecRule "ARGS:foo""@rx select""phase:2,log,auditlog,pass,id:1111,t:utf8toUnicode,t:urlDecodeUni"
T (0) Utf8toUnicode:"'%u00fan%u00ed%u00f3ns%u00e9l%u00e9ct data fr%u00f3m n%u00e1m%u00e9s"
T (0) urlDecodeUni:"'union select data from names"
Transformation completed in 29 usec.
Executing operator "rx" withparam "select" against ARGS:foo.
Target value: "'union select data fromnames"
Operator completed in 16 usec.
Warning. Pattern match "select"at ARGS:foo.
Rulereturned 1.
As you can see, the sequence "úníón+séléct+data+fróm+names" is now normalized to "union select data from names" and thus the rule matched. Summarizing the engine steps for data normalization:
- Attacker input: úníón+séléct+data+fróm+names.
- Input in UTF-8 format: \xc3\xban\xc3\xad\xc3\xb3n s\xc3\xa9l\xc3\xa9ct datafr\xc3\xb3m n\xc3\xa1m\xc3\xa9s
- Input in Unicode format: %u00fan%u00ed%u00f3n s%u00e9l%u00e9ct data fr%u00f3mn%u00e1m%u00e9s
- Input in ASCII format: union select data from names
With this new UTF-8 and Unicode mapping and decoding support, ModSecurity can now more accurately normalize data from non-english languages which results in better rule accuracy.
