The Universe of Disco


Wed, 11 Apr 2007

A security problem in a CGI program
<sarcasm>No! Who could possibly have predicted such a thing?</sarcasm>

I was away in Asia, and when I got back I noticed some oddities in my mail logs. Specifically, Yahoo! was rejecting mail.plover.com's outgoing email. In the course of investigating the mail logs, I discovered the reason why: mail.plover.com had been relaying a ton of outgoing spam.

It took me a little while to track down the problem. It was a mailing list subscription form on perl.plover.com:

your address
perl-qotw perl-qotw-discuss perl-qotw-discuss-digest

The form took the input email address and used it to manufacture an email message, requesting that that address be subscribed to the indicated lists:

        my $MAILER = '/var/qmail/bin/qmail-inject';
        ...

        for (@lists) {
          next unless m|^perl-qotw[\-a-z]*|;
          open M, "|$MAILER" or next;
          print M "From: nobody\@plover.com\n";
          print M "To: $_-subscribe-$addre\@plover.com\n";
          print M "\nRequested by $ENV{REMOTE_ADDR} at ", scalar(localtime), "\n";
          close M or next;
          push @DONE, $_;
        }
The message was delivered to the list management software, which interpreted it as a request to subscribe, and generated an appropriate confirmation reply. In theory, this doesn't open any new security holes, because a malicious remote user could also forge an identical message to the list management software without using the form.

The problem is the interpolated $addre variable. The value of this variable is essentially the address from the form. Interpolating user input into a string like this is always fraught with peril. Daniel J. Bernstein has one of the most succinct explanantions of this that I have ever seen:

The essence of user interfaces is parsing: converting an unstructured sequence of commands, in a format usually determined more by psychology than by solid engineering, into structured data.

When another programmer wants to talk to a user interface, he has to quote: convert his structured data into an unstructured sequence of commands that the parser will, he hopes, convert back into the original structured data.

This situation is a recipe for disaster. The parser often has bugs: it fails to handle some inputs according to the documented interface. The quoter often has bugs: it produces outputs that do not have the right meaning. Only on rare joyous occasions does it happen that the parser and the quoter both misinterpret the interface in the same way.

When the original data is controlled by a malicious user, many of these bugs translate into security holes.

In this case, I interpolated user data without quoting, and suffered the consequences.

The malicious remote user supplied an address of the following form:

        into9507@plover.com
        Content-Transfer-Encoding: quoted-printable
        Content-Type: text/plain
        From: Alwin Bestor <Morgan@plover.com>
        Subject: Enhance your love life with this medical marvel
        bcc: hapless@example.com,recipients@example.com,of@example.net,
             spam@example.com,...

        Attention fellow males..

        If you'd like to have stronger, harder and larger erections,
        more power and intense orgasms, increased stamina and
        ejaculatory control, and MUCH more,...

        .
(Yes, my system was used to send out penis enlargement spam. Oh, the embarrassment.)

The address contained many lines of data, separated by CRNL, and a complete message header. Interpolated into the subscription message, the bcc: line caused the qmail-inject user ineterface program to add all the "bcc" addresses to the outbound recipient list.

Several thoughts occur to me about this.

User interfaces and programmatic interfaces

The problem would probably not have occurred had I used the qmail-queue progam, which provides a programmatic interface, rather than qmail-inject, which provides a user interface. I originally selected qmail-inject for convenience: it automatically generates Date and Message-ID fields, for example. The qmail-queue program does not try to parse the recipient information from the message header; it takes recipient information in an out-of-band channel.

Perl piped open is deficient

Perl's system and exec functions have two modes. One mode looks like this:

        system "command arg arg arg...";
If the argument string contains shell metacharacters or certain other constructions, Perl uses the shell to execute the command; otherwise it forks and execs the command directly. The shell is the cause of all sorts of parsing-quoting problems, and is best avoided in programs like this one. But Perl provides an alternative:

        system "command", "arg", "arg", "arg"...;
Here Perl never uses the shell; it always forks and execs the command directly. Thus, system "cat *" prints the contents of all the files in the current working directory, but system "cat", "*" prints only the contents of the file named "*", if there is one.

qmail-inject has an option to take the envelope information from an out-of-band channel: you can supply it in the command-line arguments. I did not use this option in the original program, because I did not want to pass the user input through the Unix shell, which is what Perl's open FH, "| command args..." construction would have required.

Unfortunately, there is no easy way to avoid the shell when running a command that is attached to the parent process via a pipe. Perl provides open "| command arg arg arg...", which is what I used, and which is analogous to the first construction, involving the shell. But it provides nothing analogous to the second construction, which avoids the shell. If it did, then I probably would have used it, writing something like this:

        open M, "|", $MAILER, "-fnobody\@plover.com", $addre;
and the whole problem would have been avoided.

A better choice would have been to set up the pipe myself and use Perl's exec function to execute qmail-inject, which bypasses the shell. The qmail-inject program would always have received exactly one receipient address argument. In the event of an attack like the one above, it would have tried to send mail to into9507@plover.com^M^JContent-Transfer-Encoding:..., which would simply have bounced.

Why didn't I do this? Pure laziness.

qmail-queue more vulnerable than qmail-inject in this instance

Rather strangely, an partial attack is feasible with qmail-queue, even though it provides a (mostly) non-parsing interface. The addresses to qmail-queue are supplied to it on file descriptor 1 in the form:

        Taddress1@example.com^@Taddress2@example.com^@Taddress3@example.com^@...^@
If my program were to talk to qmail-queue instead of to qmail-inject, the program would have contained code that looked like this:

        print QMAIL_QUEUE_ENVELOPE "T$addre\0";
qmail-queue parses only to the extent of dividing up its input at the ^@ characters. But even this little bit of parsing is a problem. By supplying an appropriately-formed address string, a malicious user could still have forced my program to send mail to many addresses.

But still the recipient addresses would have been out of the content of the message. If the malicious user is unable to affect the content of the message body, the program is not useful for spamming.

But using qmail-queue, my program would have had to generate the To field itself, and so it would have had to insert the user-supplied address into the content of the message. This would have opened the whole can of worms again.

My program attacked specifically

I think some human put real time into attacking this particular program. There are bots that scour the web for email submission forms, and then try to send spam through them. Those bots don't successfully attack this program, because the recipient address is hard-wired. Also, the program refuses to send email unless at least one of the checkboxes is checked, and form-spam bots don't typically check boxes. Someone had to try some experiments to get the input format just so. I have logs of the experiments.

A couple of days after the exploit was discovered, a botnet started using it to send spam; 42 different IP addresses sent requests. I fixed the problem last night around 22:30; there were about 320 more requests, and by 09:00 this morning the attempts to send spam stopped.

Perl's "taint" feature would not have prevented this

Perl offers a feature that is designed specifically for detecting and preventing exactly this sort of problem. It tracks which data are possibly under control of a malicious user, and whether they are used in unsafe operations. Unsafe operations include most file and process operations.

One of my first thoughts was that I should have enabled the tainting feature to begin with. However, it would not have helped in this case. Although the user-supplied address would have been flagged as "tainted" and so untrustworthy; by extension, the email message string into which it was interpolated would have been tainted. But Perl does not consider writing a tainted string to a pipe to be an "unsafe" operation and so would not have signalled a failure.

The fix

The short-range fix to the problem was simple:

        my $addr = param('addr');
        $addr =~ s/^\s+//;
        $addr =~ s/\s+$//;

        ...

        if ($addr =~ /\s/) {
          sleep 45 + rand(45);
          print p("Sorry, addresses are not allowed to contain spaces");
          fin();
        }
Addresses are not allowed to contain whitespace, except leading or trailing whitespace, which is ignored. Since whitespace inside an address is unlikely to be an innocent mistake, the program waits before responding, to slow down the attacker.

Summary

Dammit.

[ Addendum 20070412: There is a followup article to this one. ]


[Other articles in category /oops] permanent link