Developing Filters

Filters are tricky. They need to:

• work with a variety of the versions of the software that generates the logs;
• work with the range of logging configuration options available in the software;
• work with multiple operating systems;
• not make assumptions about the log format in excess of the software (e.g. do not assume a username doesn’t contain spaces and use S+ unless you’ve checked the source code);
• account for how future versions of the software will log messages (e.g. guess what would happen to the log message if different authentication types are added);
• not be susceptible to DoS vulnerabilities (see Filter Security below); and
• match intended log lines only.

Please follow the steps from Filter Test Cases to Developing Filter Regular Expressions and submit a GitHub pull request (PR) afterwards. If you get stuck, you can push your unfinished changes and still submit a PR – describe what you have done, what is the hurdle, and we’ll attempt to help (PR will be automagically updated with future commits you would push to complete it).

Filter Test Cases

Purpose

Start by finding the log messages that the application generates related to some form of authentication failure. If you are adding to an existing filter think about whether the log messages are of a similar importance and purpose to the existing filter. If you were a user of Fail2Ban, and did a package update of Fail2Ban that started matching new log messages, would anything unexpected happen? Would the bantime/findtime for the jail be appropriate for the new log messages? If it doesn’t, perhaps it needs to be in a separate filter definition, for example like exim filter aims at authentication failures and exim-spam at log messages related to spam.

Even if it is a new filter you may consider separating the log messages into different filters based on purpose.

Cause

Are some of the log lines a result of the same action? For example, is a PAM failure log message, followed by an application specific failure message the result of the same user/script action? If you add regular expressions for both you would end up with two failures for a single action. Therefore, select the most appropriate log message and document the other log message) with a test case not to match it and a description as to why you chose one over another.

With the selected log lines consider what action has caused those log messages and whether they could have been generated by accident? Could the log message be occurring due to the first step towards the application asking for authentication? Could the log messages occur often? If some of these are true make a note of this in the jail.conf example that you provide.

Samples

It is important to include log file samples so any future change in the regular expression will still work with the log lines you have identified.

The sample log messages are provided in a file under testcases/files/logs/ named identically as the corresponding filter (but without .conf extension). Each log line should be preceded by a line with failJSON metadata (so the logs lines are tested in the test suite) directly above the log line. If there is any specific information about the log message, such as version or an application configuration option that is needed for the message to occur, include this in a comment (line beginning with #) above the failJSON metadata.

Log samples should include only one, definitely not more than 3, examples of log messages of the same form. If log messages are different in different versions of the application log messages that show this are encouraged.

Also attempt to inject an IP into the application (e.g. by specifying it as a username) so that Fail2Ban possibly detects the IP from user input rather than the true origin. See the Filter Security section and the top example in testcases/files/logs/apache-auth as to how to do this. One you have discovered that this is possible, correct the regex so it doesn’t match and provide this as a test case with “match”: false (see failJSON below).

If the mechanism to create the log message isn’t obvious provide a configuration and/or sample scripts testcases/files/config/{filtername} and reference these in the comments above the log line.

FailJSON metadata

A failJSON metadata is a comment immediately above the log message. It will look like:

# failJSON: { "time": "2013-06-10T10:10:59", "match": true , "host": "93.184.216.119" }

Time should match the time of the log message. It is in a specific format of Year-Month-Day’T’Hour:minute:Second. If your log message does not include a year, like the example below, the year should be listed as 2005, if before Sun Aug 14 10am UTC, and 2004 if afterwards. Here is an example failJSON line preceding a sample log line:

# failJSON: { "time": "2005-03-24T15:25:51", "match": true , "host": "198.51.100.87" }
Mar 24 15:25:51 buffalo1 dropbear[4092]: bad password attempt for 'root' from 198.51.100.87:5543

The “host” in failJSON should contain the IP or domain that should be blocked.

For long lines that you do not want to be matched (e.g. from log injection attacks) and any log lines to be excluded (see “Cause” section above), set “match”: false in the failJSON and describe the reason in the comment above.

After developing regexes, the following command will test all failJSON metadata against the log lines in all sample log files:

./fail2ban-testcases testSampleRegex

Developing Filter Regular Expressions

Date/Time

At the moment, Fail2Ban depends on log lines to have time stamps. That is why before starting to develop failregex, check if your log line format known to Fail2Ban. Copy the time component from the log line and append an IP address to test with following command:

./fail2ban-regex "2013-09-19 02:46:12 1.2.3.4" "<HOST>"

Output of such command should contain something like:

Date template hits:
|- [# of hits] date format
|  [1] Year-Month-Day Hour:Minute:Second

Ensure that the template description matches time/date elements in your log line time stamp. If there is no matched format then date template needs to be added to server/datedetector.py. Ensure that a new template is added in the order that more specific matches occur first and that there is no confusion between a Day and a Month.

Filter file

The filter is specified in a config/filter.d/{filtername}.conf file. Filter file can have sections INCLUDES (optional) and Definition as follows:

[INCLUDES]

before = common.conf

after = filtername.local

[Definition]

failregex = ....

ignoreregex = ....

This is also documented in the man page jail.conf (section 5). Other definitions can be added to make failregex’s more readable and maintainable to be used through string Interpolations (see http://docs.python.org/2.7/library/configparser.html)

General rules

Use “before” if you need to include a common set of rules, like syslog or if there is a common set of regexes for multiple filters.

Use “after” if you wish to allow the user to overwrite a set of customisations of the current filter. This file doesn’t need to exist.

Try to avoid using ignoreregex mainly for performance reasons. The case when you would use it is if in trying to avoid using it, you end up with an unreadable failregex.

Syslog

If your application logs to syslog you can take advantage of log line prefix definitions present in common.conf. So as a base use:

[INCLUDES]

before = common.conf

[Definition]

_daemon = app

failregex = ^%(__prefix_line)s

In this example common.conf defines __prefix_line which also contains the _daemon name (in syslog terms the service) you have just specified. _daemon can also be a regex.

For example, to capture following line _daemon should be set to “dovecot”:

Dec 12 11:19:11 dunnart dovecot: pop3-login: Aborted login (tried to use disabled plaintext auth): rip=190.210.136.21, lip=113.212.99.193

and then ^%(__prefix_line)s would match “Dec 12 11:19:11 dunnart dovecot: ”. Note it matches the trailing space(s) as well.

Substitutions (AKA string interpolations)

We have used string interpolations in above examples. They are useful for making the regexes more readable, reuse generic patterns in multiple failregex lines, and also to refer definition of regex parts to specific filters or even to the user. General principle is that value of a _name variable replaces occurrences of %(_name)s within the same section or anywhere in the config file if defined in [DEFAULT] section.

Regular Expressions

Regular expressions (failregex, ignoreregex) assume that the date/time has been removed from the log line (this is just how fail2ban works internally ATM).

If the format is like ‘<date...> error 1.2.3.4 is evil’ then you need to match the < at the start so regex should be similar to ‘^<> <HOST> is evil$’ using <HOST> where the IP/domain name appears in the log line.

The following general rules apply to regular expressions:

• ensure regexes start with a ^ and are as restrictive as possible. E.g. do not use .* if d+ is sufficient;
• use functionality of Python regexes defined in the standard Python re library http://docs.python.org/2/library/re.html;
• make regular expressions readable (as much as possible). E.g. (?:...) represents a non-capturing regex but (...) is more readable, thus preferred.

If you have only a basic knowledge of regular repressions we advise to read http://docs.python.org/2/library/re.html first. It doesn’t take long and would remind you e.g. which characters you need to escape and which you don’t.

Developing/testing a regex

You can develop a regex in a file or using command line depending on your preference. You can also use samples you have already created in the test cases or test them one at a time.

The general tool for testing Fail2Ban regexes is fail2ban-regex. To see how to use it run:

./fail2ban-regex --help

Take note of -l heavydebug / -l debug and -v as they might be very useful.

Tip

Take a look at the source code of the application you are developing failregex for. You may see optional or extra log messages, or parts there of, that need to form part of your regex. It may also reveal how some parts are constrained and different formats depending on configuration or less common usages.

Tip

For looking through source code - http://sourcecodebrowser.com/ . It has call graphs and can browse different versions.

Tip

Some applications log spaces at the end. If you are not sure add s*$ as the end part of the regex.

If your regex is not matching, http://www.debuggex.com/?flavor=python can help to tune it. fail2ban-regex -D ... will present Debuggex URLs for the regexs and sample log files that you pass into it.

In general use when using regex debuggers for generating fail2ban filters: * use regex from the ./fail2ban-regex output (to ensure all substitutions are done) * replace <HOST> with (?&.ipv4) * make sure that regex type set to Python * for the test data put your log output with the date/time removed

When you have fixed the regex put it back into your filter file.

Please spread the good word about Debuggex - Serge Toarca is kindly continuing its free availability to Open Source developers.

Finishing up

If you’ve added a new filter, add a new entry in config/jail.conf. The theory here is that a user will create a jail.local with [filtername]nenable=true to enable your jail.

So more specifically in the [filter] section in jail.conf:

• ensure that you have “enabled = false” (users will enable as needed);
• use “filter =” set to your filter name;
• use a typical action to disable ports associated with the application;
• set “logpath” to the usual location of application log file;
• if the default findtime or bantime isn’t appropriate to the filter, specify more appropriate choices (possibly with a brief comment line).

Submit github pull request (See “Pull Requests” above) for github.com/fail2ban/fail2ban containing your great work.

Filter Security

Poor filter regular expressions are susceptible to DoS attacks.

When a remote user has the ability to introduce text that would match filter’s failregex, while matching inserted text to the <HOST> part, they have the ability to deny any host they choose.

So the <HOST> part must be anchored on text generated by the application, and not the user, to an extent sufficient to prevent user inserting the entire text matching this or any other failregex.

Ideally filter regex should anchor at the beginning and at the end of log line. However as more applications log at the beginning than the end, anchoring the beginning is more important. If the log file used by the application is shared with other applications, like system logs, ensure the other application that use that log file do not log user generated text at the beginning of the line, or, if they do, ensure the regexes of the filter are sufficient to mitigate the risk of insertion.

Examples of poor filters

1. Too restrictive

We find a log message:

Apr-07-13 07:08:36 Invalid command fial2ban from 1.2.3.4

We make a failregex:

^Invalid command \S+ from <HOST>

Now think evil. The user does the command ‘blah from 1.2.3.44’

The program diligently logs:

Apr-07-13 07:08:36 Invalid command blah from 1.2.3.44 from 1.2.3.4

And fail2ban matches 1.2.3.44 as the IP that it ban. A DoS attack was successful.

The fix here is that the command can be anything so .* is appropriate:

^Invalid command .* from <HOST>

Here the .* will match until the end of the string. Then realise it has more to match, i.e. “from <HOST>” and go back until it find this. Then it will ban 1.2.3.4 correctly. Since the <HOST> is always at the end, end the regex with a $:

^Invalid command .* from <HOST>$

Note if we’d just had the expression:

^Invalid command \S+ from <HOST>$

Then provided the user put a space in their command they would have never been banned.

2. Unanchored regex can match other user injected data

From the Apache vulnerability CVE-2013-2178 ( original ref: https://vndh.net/note:fail2ban-089-denial-service ).

An example bad regex for Apache:

failregex = [[]client <HOST>[]] user .* not found

Since the user can do a get request on:

GET /[client%20192.168.0.1]%20user%20root%20not%20found HTTP/1.0
Host: remote.site

Now the log line will be:

[Sat Jun 01 02:17:42 2013] [error] [client 192.168.33.1] File does not exist: /srv/http/site/[client 192.168.0.1] user root not found

As this log line doesn’t match other expressions hence it matches the above regex and blocks 192.168.33.1 as a denial of service from the HTTP requester.

3. Over greedy pattern matching

From: https://github.com/fail2ban/fail2ban/pull/426

An example ssh log (simplified):

Sep 29 17:15:02 spaceman sshd[12946]: Failed password for user from 127.0.0.1 port 20000 ssh1: ruser remoteuser

As we assume username can include anything including spaces its prudent to put .* here. The remote user can also exist as anything so lets not make assumptions again:

failregex = ^%(__prefix_line)sFailed \S+ for .* from <HOST>( port \d*)?( ssh\d+)?(: ruser .*)?$

So this works. The problem is if the .* after remote user is injected by the user to be ‘from 1.2.3.4’. The resultant log line is:

Sep 29 17:15:02 spaceman sshd[12946]: Failed password for user from 127.0.0.1 port 20000 ssh1: ruser from 1.2.3.4

Testing with:

fail2ban-regex -v 'Sep 29 17:15:02 Failed password for user from 127.0.0.1 port 20000 ssh1: ruser from 1.2.3.4' '^ Failed \S+ for .* from <HOST>( port \d*)?( ssh\d+)?(: ruser .*)?$'

Tip

I’ve removed the bit that matches __prefix_line from the regex and log.

Shows:

1) [1] ^ Failed \S+ for .* from <HOST>( port \d*)?( ssh\d+)?(: ruser .*)?$
   1.2.3.4  Sun Sep 29 17:15:02 2013

It should of matched 127.0.0.1. So the first greedy part of the greedy regex matched until the end of the string. The was no “from <HOST>” so the regex engine worked backwards from the end of the string until this was matched.

The result was that 1.2.3.4 was matched, injected by the user, and the wrong IP was banned.

The solution here is to make the first .* non-greedy with .*?. Here it matches as little as required and the fail2ban-regex tool shows the output:

fail2ban-regex -v 'Sep 29 17:15:02 Failed password for user from 127.0.0.1 port 20000 ssh1: ruser from 1.2.3.4' '^ Failed \S+ for .*? from <HOST>( port \d*)?( ssh\d+)?(: ruser .*)?$'

1) [1] ^ Failed \S+ for .*? from <HOST>( port \d*)?( ssh\d+)?(: ruser .*)?$
   127.0.0.1  Sun Sep 29 17:15:02 2013

So the general case here is a log line that contains:

(fixed_data_1)<HOST>(fixed_data_2)(user_injectable_data)

Where the regex that matches fixed_data_1 is gready and matches the entire string, before moving backwards and user_injectable_data can match the entire string.

Another case

ref: https://www.debuggex.com/r/CtAbeKMa2sDBEfA2/0

A webserver logs the following without URL escaping:

[error] 2865#0: *66647 user "xyz" was not found in "/file", client: 1.2.3.1, server: www.host.com, request: "GET ", client: 3.2.1.1, server: fake.com, request: "GET exploited HTTP/3.3", host: "injected.host", host: "www.myhost.com"

regex:

failregex = ^ \[error\] \d+#\d+: \*\d+ user "\S+":? (?:password mismatch|was not found in ".*"), client: <HOST>, server: \S+, request: "\S+ .+ HTTP/\d+\.\d+", host: "\S+"

The .* matches to the end of the string. Finds that it can’t continue to match ”, client ... so it moves from the back and find that the user injected web URL:

", client: 3.2.1.1, server: fake.com, request: "GET exploited HTTP/3.3", host: "injected.host

In this case there is a fixed host: “www.myhost.com” at the end so the solution is to anchor the regex at the end with a $.

If this wasn’t the case then first .* needed to be made so it didn’t capture beyond <HOST>.

4. Application generates two identical log messages with different meanings

If the application generates the following two messages under different circumstances:

client <IP>: authentication failed
client <USER>: authentication failed

Then it’s obvious that a regex of ^client <HOST>: authentication failed$ will still cause problems if the user can trigger the second log message with a <USER> of 123.1.1.1.

Here there’s nothing to do except request/change the application so it logs messages differently.