I recently became aware of Regular expression Denial of Service attacks, and decided to root out so-called 'evil' regex patterns wherever I could find them in my codebase - or at least those that are used on user input. The examples given at the OWASP link above and wikipedia are helpful, but they don't do a great job of explaining the problem in simple terms.
A description of evil regexes, from wikipedia:
the regular expression applies repetition ("+", "*") to a complex subexpression;
for the repeated subexpression, there exists a match which is also a suffix of another valid match.
With examples, again from wikipedia:
(a+)+
([a-zA-Z]+)*
(a|aa)+
(a|a?)+
(.*a){x} for x > 10
Is this a problem that just doesn't have a simpler explanation? I'm looking for something that would make it easier to avoid this problem while writing regexes, or to find them within an existing codebase.
Why Are Evil Regexes A Problem?
Because computers do exactly what you tell them to do, even if it's not what you meant or is totally unreasonable. If you ask a regex engine to prove that, for some given input, there either is or is not a match for a given pattern, then the engine will attempt to do that no matter how many different combinations must be tested.
Here is a simple pattern inspired by the first example in the OP's post:
^((ab)*)+$
Given the input:
abababababababababababab
The regex engine tries something like (abababababababababababab) and a match is found on the first try.
But then we throw the monkey wrench in:
abababababababababababab a
The engine will first try (abababababababababababab) but that fails because of that extra a. This causes catastrophic backtracking, because our pattern (ab)*, in a show of good faith, will release one of its captures (it will "backtrack") and let the outer pattern try again. For our regex engine, that looks something like this:
(abababababababababababab) - Nope
(ababababababababababab)(ab) - Nope
(abababababababababab)(abab) - Nope
(abababababababababab)(ab)(ab) - Nope
(ababababababababab)(ababab) - Nope
(ababababababababab)(abab)(ab) - Nope
(ababababababababab)(ab)(abab) - Nope
(ababababababababab)(ab)(ab)(ab) - Nope
(abababababababab)(abababab) - Nope
(abababababababab)(ababab)(ab) - Nope
(abababababababab)(abab)(abab) - Nope
(abababababababab)(abab)(ab)(ab) - Nope
(abababababababab)(ab)(ababab) - Nope
(abababababababab)(ab)(abab)(ab) - Nope
(abababababababab)(ab)(ab)(abab) - Nope
(abababababababab)(ab)(ab)(ab)(ab) - Nope
(ababababababab)(ababababab) - Nope
(ababababababab)(abababab)(ab) - Nope
(ababababababab)(ababab)(abab) - Nope
(ababababababab)(ababab)(ab)(ab) - Nope
(ababababababab)(abab)(abab)(ab) - Nope
(ababababababab)(abab)(ab)(abab) - Nope
(ababababababab)(abab)(ab)(ab)(ab) - Nope
(ababababababab)(ab)(abababab) - Nope
(ababababababab)(ab)(ababab)(ab) - Nope
(ababababababab)(ab)(abab)(abab) - Nope
(ababababababab)(ab)(abab)(ab)(ab) - Nope
(ababababababab)(ab)(ab)(ababab) - Nope
(ababababababab)(ab)(ab)(abab)(ab) - Nope
(ababababababab)(ab)(ab)(ab)(abab) - Nope
(ababababababab)(ab)(ab)(ab)(ab)(ab) - Nope
...
(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(abababab) - Nope
(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ababab)(ab) - Nope
(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(abab)(abab) - Nope
(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(abab)(ab)(ab) - Nope
(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ababab) - Nope
(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(abab)(ab) - Nope
(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(abab) - Nope
(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab)(ab) - Nope
The number of possible combinations scales exponentially with the length of the input and, before you know it, the regex engine is eating up all your system resources trying to solve this thing until, having exhausted every possible combination of terms, it finally gives up and reports "There is no match." Meanwhile your server has turned into a burning pile of molten metal.
How to Spot Evil Regexes
It's actually very tricky. Catastrophic backtracking in modern regex engines is similar in nature to the halting problem which Alan Turing proved was impossible to solve. I have written problematic regexes myself, even though I know what they are and generally how to avoid them. Wrapping everything you can in an atomic group can help to prevent the backtracking issue. It basically tells the regex engine not to revisit a given expression - "lock whatever you matched on the first try". Note, however, that atomic expressions don't prevent backtracking within the expression, so ^(?>((ab)*)+)$ is still dangerous, but ^(?>(ab)*)+$ is safe (it'll match (abababababababababababab) and then refuse to give up any of it's matched characters, thus preventing catastrophic backtracking).
Unfortunately, once it's written, it's actually very hard to immediately or quickly find a problem regex. In the end, recognizing a bad regex is like recognizing any other bad code - it takes a lot of time and experience and/or a single catastrophic event.
Interestingly, since this answer was first written, a team at the University of Texas at Austin published a paper describing the development of a tool capable of performing static analysis of regular expressions with the express purpose of finding these "evil" patterns. The tool was developed to analyse Java programs, but I suspect that in the coming years we'll see more tools developed around analysing and detecting problematic patterns in JavaScript and other languages, especially as the rate of ReDoS attacks continues to climb.
Static Detection of DoS Vulnerabilities in
Programs that use Regular Expressions
Valentin Wüstholz, Oswaldo Olivo, Marijn J. H. Heule, and Isil Dillig
The University of Texas at Austin
Detecting evil regexes
Try Nicolaas Weideman's RegexStaticAnalysis project.
Try my ensemble-style vuln-regex-detector which has a CLI for Weideman's tool and others.
Rules of thumb
Evil regexes are always due to ambiguity in the corresponding NFA, which you can visualize with tools like regexper.
Here are some forms of ambiguity. Don't use these in your regexes.
Nesting quantifiers like (a+)+ (aka "star height > 1"). This can cause exponential blow-up. See substack's safe-regex tool.
Quantified Overlapping Disjunctions like (a|a)+. This can cause exponential blow-up.
Avoid Quantified Overlapping Adjacencies like \d+\d+. This can cause polynomial blow-up.
Additional resources
I wrote this paper on super-linear regexes. It includes loads of references to other regex-related research.
What you call an "evil" regex is a regex that exhibits catastrophic backtracking. The linked page (which I wrote) explains the concept in detail. Basically, catastrophic backtracking happens when a regex fails to match and different permutations of the same regex can find a partial match. The regex engine then tries all those permutations. If you want to go over your code and inspect your regexes these are the 3 key issues to look at:
Alternatives must be mutually exclusive. If multiple alternatives can match the same text then the engine will try both if the remainder of the regex fails. If the alternatives are in a group that is repeated, you have catastrophic backtracking. A classic example is (.|\s)* to match any amount of any text when the regex flavor does not have a "dot matches line breaks" mode. If this is part of a longer regex then a subject string with a sufficiently long run of spaces (matched by both . and \s) will break the regex. The fix is to use (.|\n)* to make the alternatives mutually exclusive or even better to be more specific about which characters are really allowed, such as [\r\n\t\x20-\x7E] for ASCII printables, tabs, and line breaks.
Quantified tokens that are in sequence must either be mutually exclusive with each other or be mutually exclusive what comes between them. Otherwise both can match the same text and all combinations of the two quantifiers will be tried when the remainder of the regex fails to match. A classic example is a.*?b.*?c to match 3 things with "anything" between them. When c can't be matched the first .*? will expand character by character until the end of the line or file. For each expansion the second .*? will expand character by character to match the remainder of the line or file. The fix is to realize that you can't have "anything" between them. The first run needs to stop at b and the second run needs to stop at c. With single characters a[^b]*+b[^c]*+c is an easy solution. Since we now stop at the delimiter, we can use possessive quantifiers to further increase performance.
A group that contains a token with a quantifier must not have a quantifier of its own unless the quantified token inside the group can only be matched with something else that is mutually exclusive with it. That ensures that there is no way that fewer iterations of the outer quantifier with more iterations of the inner quantifier can match the same text as more iterations of the outer quantifier with fewer iterations of the inner quantifier. This is the problem illustrated in JDB's answer.
While I was writing my answer I decided that this merited a full article on my website. This is now online too.
I would sum it up as "A repetition of a repetition". The first example you listed is a good one, as it states "the letter a, one or more times in a row. This can again happen one or more times in a row".
What to look for in this case is combination of the quantifiers, such as * and +.
A somewhat more subtle thing to look out for is the third and fourth one. Those examples contain an OR operation, in which both sides can be true. This combined with a quantifier of the expression can result in a LOT of potential matches depending on the input string.
To sum it up, TLDR-style:
Be careful how quantifiers are used in combination with other operators.
I have surprisingly come across ReDOS quite a few times performing source code reviews. One thing I would recommend is to use a timeout with whatever Regular Expression engine that you are using.
For example, in C# I can create the regular expression with a TimeSpan attribute.
string pattern = #"^<([a-z]+)([^<]+)*(?:>(.*)<\/\1>|\s+\/>)$";
Regex regexTags = new Regex(pattern, RegexOptions.None, TimeSpan.FromSeconds(1.0));
try
{
string noTags = regexTags.Replace(description, "");
System.Console.WriteLine(noTags);
}
catch (RegexMatchTimeoutException ex)
{
System.Console.WriteLine("RegEx match timeout");
}
This regex is vulnerable to denial of service and without the timeout will spin and eat resources. With the timeout, it will throw a RegexMatchTimeoutException after the given timeout and will not cause the resource usage leading to a Denial of Service condition.
You will want to experiment with the timeout value to make sure it works for your usage.
I would say this is related to the regex engine in use. You may not always be able to avoid these types of regexes, but if your regex engine is built right, then it is less of a problem. See this blog series for a great deal of information on the topic of regex engines.
Note the caveat at the bottom of the article, in that backtracking is an NP-Complete problem. There currently is no way to efficiently process them, and you might want to disallow them in your input.
I don't think you can recognize such regexes, at least not all of them or not without restrictively limiting their expressiveness. If you'd really care about ReDoSs, I'd try to sandbox them and kill their processing with a timeout. It also might be possible that there are RegEx implementations that let you limit their max backtracking amount.
There are some ways I can think of that you could implement some simplification rules by running them on small test inputs or analyzing the regex's structure.
(a+)+ can be reduced using some sort of rule for replacing redundant operators to just (a+)
([a-zA-Z]+)* could also be simplified with our new redundancy combining rule to ([a-zA-Z]*)
The computer could run tests by running the small subexpressions of the regex against randomly-generated sequences of the relevant characters or sequences of characters, and seeing what groups they all end up in. For the first one, the computer is like, hey the regex wants a's, so lets try it with 6aaaxaaq. It then sees that all the a's, and only the first groupm end up in one group, and concludes that no matter how many a's is puts, it won't matter, since + gets all in the group. The second one, is like, hey, the regex wants a bunch of letters, so lets try it with -fg0uj=, and then it sees that again each bunch is all in one group, so it gets rid of the + at the end.
Now we need a new rule to handle the next ones: The eliminate-irrelevant-options rule.
With (a|aa)+, the computer takes a look at it and is like, we like that big second one, but we can use that first one to fill in more gaps, lets get ans many aa's as we can, and see if we can get anything else after we're done. It could run it against another test string, like `eaaa#a~aa.' to determine that.
You can protect yourself from (a|a?)+ by having the computer realize that the strings matched by a? are not the droids we are looking for, because since it can always match anywhere, we decide that we don't like things like (a?)+, and throw it out.
We protect from (.*a){x} by getting it to realize that the characters matched by a would have already been grabbed by .*. We then throw out that part and use another rule to replace the redundant quantifiers in (.*){x}.
While implementing a system like this would be very complicated, this is a complicated problem, and a complicated solution may be necessary. You should also use techniques other people have brought up, like only allowing the regex some limited amount of execution resources before killing it if it doesn't finish.
Related
Is there an algorithm to determine whether a given JavaScript regex is vulnerable to ReDoS? The algorithm doesn't have to be perfect - some false positives and false negatives are acceptable. (I'm specifically interested in ECMA-262 regexes.)
It is hard to verify whether a regexp is evil or not without actually running it. You could try detecting some of the patterns detailed in the Wiki and generalise them:
e.g. For
(a+)+
([a-zA-Z]+)*
(a|aa)+
(a|a?)+
(.*a){x} for x > 10
You could check for )+ or )* or ){ sequences and validate against them. However, I guarantee that an attacker will find their way round them.
In essence it is a minefield to allow user set regexps. However, if you can timeout the regexp search, terminate the thread and then mark that regexp as "bad" you can mitigate the threat somewhat. In the case that the regexp is used later, maybe you could validate it by running it against an expected input at point of entry?
Later you will still need to be able to terminate it if the text evaluated at the later stage has a different effect with your regexp and mark it as bad so it will not be used again without user intervention.
TL;DR sort of, but not fully
In [9]: re.compile("(a+)+", re.DEBUG)
max_repeat 1 4294967295
subpattern 1
max_repeat 1 4294967295
literal 97
Note those nested repeat 1..N, for large N, that's bad.
This takes care of all Wikipedia examples, except (a|aa)+and a*b?a*x.
Likewise it's hard to account for back-references, if your engine supports those.
IMO evil regexp is combination of two factors: combinatorial explosion and oversight in engine implementation. Thus, worst case also depends on regexp engine and sometimes flags. Backtracking is not always easy to identify.
Simple cases, however, can be identified.
Disclaimer: my question is not focused on the exercise, it's just an example (although if you have any interesting tips on the example itself, feel free to share!).
Say I'm working with parsing some strings with Regex in JavaScript, and my main focus is performance (speed).
I have a piece of regex which checks for a numeric string, and then parses it using Number if it's numeric:
if (/^\[[0-9]+]$/.test(str)) {
val = Number(str.match(/^\[([0-9]+)$/)[1]);
}
Note how the conditional test does not have a capture group around the digits. This leads to writing out basically the same regex twice, except with a capture group the second time.
What I would like to know is this; does adding a capture group to a regex used alongside test() in a condition affect performance in any way? I'd like to simply use the capture regex in both places, as long as there is no performance hit.
And to the question as why I'm doing test() then match() rather than match() and checking null; I want to keep parsing as fast as possible when there's a miss, but it's ok to be a little slower when there's a hit.
If it's not clear from the above, I'm referring to JavaScript's regex engine - although if this differs across engines it'd be nice to know too. I'm working specifically in Node.js here, should it also differ across JS engines.
Thanks in advance!
Doing 2 regexps - that are very similar in scope - will almost always be slower than doing a single one because regexps are greedy (that means that they will try to match as much as they can, usually meaning take the maximum amount of time possible).
What you're asking is basically: is the cost of fewer memory in the worst case scenario (aka using the .test to save on memory from capture) faster than just using the extra memory? The answer is no, using extra memory speeds up your process.
Don't take my word for it though, here's a jsperf: http://jsperf.com/regex-perf-numbers
I have the following regular expression that works fine in my application code and other code editors have not reported a problem with it. It is used to validate a password.
/^(?=.*[A-Za-z])+(?=.*[\d])+(?=.*[^A-Za-z\d\s])+.*$/
So in other words:
Must have one letter
Must have one digit
Must have one non-letter, non-digit
Now it seems netbeans has a fairly decent regex parser and it has reported that this is an erroneous statement. But as i am new to regex I cannot spot the error. Is it due to using the positive lookahead ?= with the one or more + at the end?
When I take out the + the error goes away, but the regex stops performing in my application.
If anyone can tell me what is wrong with my expression that would be great.
The statement is used in a jQuery validation plugin that i use, if that helps. Also due to the fact I am using a plugin, I would prefer not splitting this into several smaller (clearly simpler and cleaner) expressions. That would require a great deal of work.
It never makes sense to apply a quantifier to a zero-width assertion such as a lookahead. The whole point of such assertions is that they allow you to assert that some condition is true, without consuming any of the text--that is, advancing the current match position. Some regex flavors treat that as a syntax error, while others effectively ignore the quantifier. Getting rid of those plus signs makes your regex correct:
/^(?=.*[A-Za-z])(?=.*\d)(?=.*[^A-Za-z\d\s]).*$/
If it doesn't work as expected, you may be running into the infamous IE lookahead bug. The usual workaround is to reorder things so the first lookahead is anchored at the end, like so:
/^(?=.{8,15}$)(?=.*[A-Za-z])(?=.*\d)(?=.*[^A-Za-z\d\s]).*/
The (?=.{8,15}$) is just an example; I have no idea what your real requirements are. If you do want to impose minimum and maximum length limits, this is the ideal place to do it.
I was writing some regexes in my text editor (Sublime) today in an attempt to quickly find specific segments of source code, and it required getting a little creative because sometimes the function call might contain more function calls. For example I was looking for jQuery selectors:
$("div[class='should_be_using_dot_notation']");
$(escapeJQSelector("[name='crazy{"+getName(object)+"}']"));
I don't consider it unreasonable to expect one of my favorite powertools (regex) to help me do this sort of searching, but it's clear that the expression required to parse the second bit of code there will be somewhat complex as there are two levels of nested parens.
I am sufficiently versed in the theory to know that this sort of parsing is exactly what a context-free grammar parser is for, and that building out a regex is likely to suck up more memory and time (perhaps in an exponential rather than O(n^3) fashion). However I am not expecting to see that sort of feature available in my text editor or web browser any time soon, and I just wanted to squeak by with a big nasty regex.
Starting from this (This matches zero levels of nested parens, and no trivial empty ones):
\$\([^)(]+?\)
Here's what the one-level nested parens one I came up with looks like:
\$\(((\([^)(]*\))|[^)(])+?\)
Breaking it down:
\$\( begin text
( groups the contents of the $() call
(\( groups a level 1 nested pair of parens
[^)(]* only accept a valid pair of parens (it shall contain anything but parens)
\)) close level 1 nesting
| contents also can be
[^)(] anything else that also is not made of parens
)+? not sure if this should be plus or star or if can be greedy (the contents are made up of either a level 1 paren group or any other character)
\) end
This worked great! But I need one more level of nesting.
I started typing up the two-level nested expression in my editor and it began to pause for 2-3 seconds at a time when I put in *'s.
So I gave up on that and moved to regextester.com, and before very long at all, the entire browser tab was frozen.
My question is two-fold.
What's a good way of constructing an arbitrary-level regex? Is this something that only human pattern-recognition can ever hope to achieve? It seems to me that I can get a good deal of intuition for how to go about making the regex capable of matching two levels of nesting based on the similarities between the first two. I think this could just be distilled down into a few "guidelines".
Why does regex parsing on non-enormous regexes block or freeze for so long?
I understand the O(n) linear time is for n where n is length of input to run the regex over (i.e. my test strings). But in a system where it recompiles the regex each time I type a new character into it, what would cause it to freeze up? Is this necessarily a bug in the regex code (I hope not, I thought the Javascript regex impl was pretty solid)? Part of my reasoning moving to a different regex tester from my editor was that I'd no longer be running it (on each keypress) over all ~2000 lines of source code, but it did not prevent the whole environment from locking up as I edited my regex. It would make sense if each character changed in the regex would correspond to some simple transformation in the DFA that represents that expression. But this appears not to be the case. If there are certain exponential time or space consequences to adding a star in a regex, it could explain this super-slow-to-update behavior.
Meanwhile I'll just go work out the next higher nested regexes by hand and copy them in to the fields once i'm ready to test them...
Um. Okay, so nobody wants to write the answer, but basically the answer here is
Backtracking
It can cause exponential runtime when you do certain non-greedy things.
The answer to the first part of my question:
The two-nested expression is as follows:
\$\(((\(((\([^)(]*\))|[^)(])*\))|[^)(])*\)
The transformation to make the next nested expression is to replace instances of [^)(]* with ((\([^)(]*\))|[^)(])*, or, as a meta-regex (where the replace-with section does not need escaping):
s/\[^\)\(\]\*/((\([^)(]*\))|[^)(])*/
This is conceptually straightforward: In the expression matching N levels of nesting, if we replace the part that forbids more nesting with something that matches one more level of nesting then we get the expression for N+1 levels of nesting!
To match an arbitrary number of nested (), with only one pair on each level of nesting, you could use the following, changing 2 to whatever number of nested () you require
/(?:\([^)(]*){2}(?:[^)(]*\)){2}/
To avoid excessive backtracking you want to avoid using nested quantifiers, particularly when the sub-pattern on both sides of an inner alternation is capable of matching the same substring.
I had a search and found lot's of similar regex examples, but not quite what I need.
I want to be able to pass in the following urls and return the results:
www.google.com returns google.com
sub.domains.are.cool.google.com returns google.com
doesntmatterhowlongasubdomainis.idont.wantit.google.com
returns google.com
sub.domain.google.com/no/thanks returns google.com
Hope that makes sense :)
Thanks in advance!-James
You can't do this with a regular expression because you don't know how many blocks are in the suffix.
For example google.com has a suffix of com. To get from subdomain.google.com to google.com you'd have to take the last two blocks - one for the suffix and one for google.
If you apply this logic to subdomain.google.co.uk though you would end up with co.uk.
You will actually need to look up the suffix from a list like http://publicsuffix.org/
Don't use regex, use the .split() method and work from there.
var s = domain.split('.');
If your use case is fairly narrow you could then check the TLDs as needed, and then return the last 2 or 3 segments as appropriate:
return s.slice(-2).join('.');
It'll make your eyes bleed less than any regex solution.
I've not done a lot of testing on this, but if I understand what you're asking for, this should be a decent starting point...
([A-Za-z0-9-]+\.([A-Za-z]{3,}|[A-Za-z]{2}\.[A-Za-z]{2}|[A-za-z]{2}))\b
EDIT:
To clarify, it's looking for:
one or more alpha-numeric characters or dashes, followed by a literal dot
and then one of three things...
three or more alpha characters (i.e. com/net/mil/coop, etc.)
two alpha characters, followed by a literal dot, followed by two more alphas (i.e. co.uk)
two alpha characters (i.e. us/uk/to, etc)
and at the end of that, a word boundary (\b) meaning the end of the string, a space, or a non-word character (in regex word characters are typically alpha-numerics, and underscore).
As I say, I didn't do much testing, but it seemed a reasonable jumping off point. You'd likely need to try it and tune it some, and even then, it's unlikely that you'll get 100% for all test cases. There are considerations like Unicode domain names and all sorts of technically-valid-but-you'll-likely-not-encounter-in-the-wild things that'll trip up a simple regex like this, but this'll probably get you 90%+ of the way there.
If you have limited subset of data, I suggest to keep the regex simple, e.g.
(([a-z\-]+)(?:\.com|\.fr|\.co.uk))
This will match:
www.google.com --> google.com
www.google.co.uk --> google.co.uk
www.foo-bar.com --> foo-bar.com
In my case, I know that all relevant URLs will be matched using this regex.
Collect a sample dataset and test it against your regex. While prototyping, you can do that using a tool such https://regex101.com/r/aG9uT0/1. In development, automate it using a test script.
([A-Za-z0-9-]+\.([A-Za-z]{3,}|[A-Za-z]{2}\.[A-Za-z]{2}|[A-za-z]{2}))(?!\.([A-Za-z]{3,}|[A-Za-z]{2}\.[A-Za-z]{2}|[A-za-z]{2}))\b
This is an improvement upon theracoonbear's answer.
I did a quick bit of testing and noticed that if you give it a domain where the subdomain has a subdomain, it will fail. I also wanted to point out that the "90%" was definitely not generous. It will be a lot closer to 100% than you think. It works on all subdomains of the top 50 most visited websites which accounts for a huge chunk of worldwide internet activity. The only time it would fail is potentially with unicode domains, etc.
My solution starts off working the same way that theracoonbear's does. Instead of checking for a word boundary, it uses a negative lookahead to check if there is not something that could be a TLD at the end (just copied the TLD checking part over into a negative lookahead).
Without testing the validity of top level domain, I'm using an adaptation of stormsweeper's solution:
domain = 'sub.domains.are.cool.google.com'
s = domain.split('.')
tld = s.slice(-2..-1).join('.')
EDIT: Be careful of issues with three part TLDs like domain.co.uk.