Network Working Group G. Ziemba
Request for Comments: 1858 Alantec
Category: Informational D. Reed
October 1995 Security Considerations for IP Fragment Filtering
Status of This Memo
This memo provides information for the Internet community. This memo
does not specify an Internet standard of any kind. Distribution of
this memo is unlimited.
IP fragmentation can be used to disguise TCP packets from IP filters
used in routers and hosts. This document describes two methods of
attack as well as remedies to prevent them.
System administrators rely on manufacturers of networking equipment
to provide them with packet filters; these filters are used for
keeping attackers from accessing private systems and information,
while permitting friendly agents to transfer data between private
nets and the Internet. For this reason, it is important for network
equipment vendors to anticipate possible attacks against their
equipment and to implement robust mechanisms to deflect such attacks.
The growth of the global Internet has brought with it an increase in
"undesirable elements" manifested in antisocial behavior. Recent
months have seen the use of novel attacks on Internet hosts, which
have in some cases led to the compromise of sensitive data.
Increasingly sophisticated attackers have begun to exploit the more
subtle aspects of the Internet Protocol; fragmentation of IP packets,
an important feature in heterogeneous internetworks, poses several
potential problems which we explore here.
2. Filtering IP Fragments
IP packet filters on routers are designed with a user interface that
hides packet fragmentation from the administrator; conceptually, an
IP filter is applied to each IP packet as a complete entity.
One approach to fragment filtering, described by Mogul , involves
keeping track of the results of applying filter rules to the first
fragment (FO==0) and applying them to subsequent fragments of the
same packet. The filtering module would maintain a list of packets
indexed by the source address, destination address, protocol, and IP
ID. When the initial (FO==0) fragment is seen, if the MF bit is set,
a list item would be allocated to hold the result of filter access
checks. When packets with a non-zero FO come in, look up the list
element with a matching SA/DA/PROT/ID and apply the stored result
(pass or block). When a fragment with a zero MF bit is seen, free
the list element.
Although this method (or some refinement of it) might successfully
remove any trace of the offending whole packet, it has some
difficulties. Fragments that arrive out of order, possibly because
they traveled over different paths, violate one of the design
assumptions, and undesired fragments can leak through as a result.
Furthermore, if the filtering router lies on one of several parallel
paths, the filtering module will not see every fragment and cannot
guarantee complete fragment filtering in the case of packets that
should be dropped.
Fortunately, we do not need to remove all fragments of an offending
packet. Since "interesting" packet information is contained in the
headers at the beginning, filters are generally applied only to the
first fragment. Non-first fragments are passed without filtering,
because it will be impossible for the destination host to complete
reassembly of the packet if the first fragment is missing, and
therefore the entire packet will be discarded.
The Internet Protocol allows fragmentation of packets into pieces so
small as to be impractical because of data and computational
overhead. Attackers can sometimes exploit typical filter behavior
and the ability to create peculiar fragment sequences in order to
sneak otherwise disallowed packets past the filter. In normal
practice, such pathalogical fragmentation is never used, so it is
safe to drop these fragments without danger of preventing normal
3. Tiny Fragment Attack
With many IP implementations it is possible to impose an unusually
small fragment size on outgoing packets. If the fragment size is
made small enough to force some of a TCP packet's TCP header fields
into the second fragment, filter rules that specify patterns for
those fields will not match. If the filtering implementation does
not enforce a minimum fragment size, a disallowed packet might be
passed because it didn't hit a match in the filter.
STD 5, RFC 791 states:
Every internet module must be able to forward a datagram of 68
octets without further fragmentation. This is because an internet
header may be up to 60 octets, and the minimum fragment is 8
Note that, for the purpose of security, it is not sufficient to
merely guarantee that a fragment contains at least 8 octets of data
beyond the IP header because important transport header information
(e.g., the CODE field of the TCP header) might be beyond the 8th data
3.1 Example of the Tiny Fragment Attack
In this example, the first fragment contains only eight octets of
data (the minimum fragment size). In the case of TCP, this is
sufficient to contain the source and destination port numbers, but
it will force the TCP flags field into the second fragment.
Filters that attempt to drop connection requests (TCP datagrams
having SYN=1 and ACK=0) will be unable to test these flags in the
first octet, and will typically ignore them in subsequent
+-+-+-+ +-+-+-+-+-+-+-+-+-+-+-+ +-+-+-+
| | ... | Fragment Offset = 0 | ... | |
+-+-+-+ +-+-+-+-+-+-+-+-+-+-+-+ +-+-+-+
| Source Port | Destination Port |
| Sequence Number |
+-+-+-+ +-+-+-+-+-+-+-+-+-+-+-+ +-+-+-+
| | ... | Fragment Offset = 1 | ... | |
+-+-+-+ +-+-+-+-+-+-+-+-+-+-+-+ +-+-+-+
| Acknowledgment Number |
| Data | |U|A|P|R|S|F| |
| Offset| Reserved |R|C|S|S|Y|I| Window |
| | |G|K|H|T|N|N| |
3.2 Prevention of the Tiny Fragment Attack
In a router, one can prevent this sort of attack by enforcing
certain limits on fragments passing through, namely, that the
first fragment be large enough to contain all the necessary header
There are two ways to guarantee that the first fragment of a
"passed" packet includes all the required fields, one direct, the
3.2.1 Direct Method
There is some number TMIN which is the minimum length of a
transport header required to contain "interesting" fields
(i.e., fields whose values are significant to packet filters).
This length is measured from the beginning of the transport
header in the original unfragmented IP packet.
Note that TMIN is a function of the transport protocol involved
and also of the particular filters currently configured.
The direct method involves computing the length of the
transport header in each zero-offset fragment and comparing it
against TMIN. If the transport header length is less than
TMIN, the fragment is discarded. Non-zero-offset fragments
need not be checked because if the zero-offset fragment is
discarded, the destination host will be unable to complete
reassembly. So far we have:
if FO=0 and TRANSPORTLEN < tmin then
However, the "interesting" fields of the common transport
protocols, except TCP, lie in the first eight octets of the
transport header, so it isn't possible to push them into a
non-zero-offset fragment. Therefore, as of this writing, only
TCP packets are vulnerable to tiny-fragment attacks and the
test need not be applied to IP packets carrying other transport
protocols. A better version of the tiny fragment test might
if FO=0 and PROTOCOL=TCP and TRANSPORTLEN < tmin then
As discussed in the section on overlapping fragments below,
however, this test does not block all fragmentation attacks,
and is in fact unnecessary when a more general technique is
3.2.2 Indirect Method
The indirect method relies on the observation that when a TCP
packet is fragmented so as to force "interesting" header fields
out of the zero-offset fragment, there must exist a fragment
with FO equal to 1.
If a packet with FO==1 is seen, conversely, it could indicate
the presence, in the fragment set, of a zero-offset fragment
with a transport header length of eight octets Discarding this
one-offset fragment will block reassembly at the receiving host
and be as effective as the direct method described above.
4. Overlapping Fragment Attack
RFC 791, the current IP protocol specification, describes a
reassembly algorithm that results in new fragments overwriting any
overlapped portions of previously-received fragments.
Given such a reassembly implementation, an attacker could construct a
series of packets in which the lowest (zero-offset) fragment would
contain innocuous data (and thereby be passed by administrative
packet filters), and in which some subsequent packet having a non-
zero offset would overlap TCP header information (destination port,
for instance) and cause it to be modified. The second packet would
be passed through most filter implementations because it does not
have a zero fragment offset.
RFC 815 outlines an improved datagram reassembly algorithm, but it
concerns itself primarily with filling gaps during the reassembly
process. This RFC remains mute on the issue of overlapping
Thus, fully-compliant IP implementations are not guaranteed to be
immune to overlapping-fragment attacks. The 4.3 BSD reassembly
implementation takes care to avoid these attacks by forcing data from
lower-offset fragments to take precedence over data from higher-
offset fragments. However, not all IP implementations are based on
the original BSD code, and it is likely that some of them are
4.1 Example of the Overlapping Fragment Attack
In this example, fragments are large enough to satisfy the minimum
size requirements described in the previous section. The filter
is configured to drop TCP connection request packets.
The first fragment contains values, e.g., SYN=0, ACK=1, that
enable it to pass through the filter unharmed.
The second fragment, with a fragment offset of eight octets,
contains TCP Flags that differ from those given in the first
fragment, e.g., SYN=1, ACK=0. Since this second fragment is not a
0-offset fragment, it will not be checked, and it, too will pass
through the filter.
The receiving host, if it conforms fully to the algorithms given
in RFC 791, will reconstitute the packet as a connection request
because the "bad" data arrived later.
If the receiving host has a reassembly algorithm that prevents new
data from overwriting data received previously, we can send
Fragment 2 first, followed by Fragment 1, and accomplish the same
4.2 Prevention of the Overlapping Fragment Attack
Since no standard requires that an overlap-safe reassembly
algorithm be used, the potential vulnerability of hosts to this
attack is quite large.
By adopting a better strategy in a router's IP filtering code, one
can be assured of blocking this "attack". If the router's
filtering module enforces a minimum fragment offset for fragments
that have non-zero offsets, it can prevent overlaps in filter
parameter regions of the transport headers.
In the case of TCP, this minimum is sixteen octets, to ensure that
the TCP flags field is never contained in a non-zero-offset
fragment. If a TCP fragment has FO==1, it should be discarded
because it starts only eight octets into the transport header.
Conveniently, dropping FO==1 fragments also protects against the
tiny fragment attack, as discussed earlier.
RFC 791 demands that an IP stack must be capable of passing an 8
byte IP data payload without further fragmentation (fragments sit
on 8 byte boundaries). Since an IP header can be up to 60 bytes
long (including options), this means that the minimum MTU on a
link should be 68 bytes.
A typical IP header is only 20 bytes long and can therefore carry
48 bytes of data. No one in the real world should EVER be
generating a TCP packet with FO=1, as it would require both that a
previous system fragmenting IP data down to the 8 byte minimum and
a 60 byte IP header.
A general algorithm, then, for ensuring that filters work in the
face of both the tiny fragment attack and the overlapping fragment
IF FO=1 and PROTOCOL=TCP then
If filtering based on fields in other transport protocol headers
is provided in a router, the minimum could be greater, depending
on the position of those fields in the header. In particular, if
filtering is permitted on data beyond the sixteenth octet of the
transport header, either because of a flexible user interface or
the implementation of filters for some new transport protocol,
dropping packets with FO==1 might not be sufficient.
5. Security Considerations
This memo is concerned entirely with the security implications of
filtering fragmented IP packets.
The attack scenarios described above grew from discussions that took
place on the firewalls mailing list during May of 1995. Participants
included: Darren Reed <email@example.com>, Tom Fitzgerald
<firstname.lastname@example.org>, and Paul Traina <email@example.com>.
 Mogul, J., "Simple and Flexible Datagram Access Controls for
Unix-based Gateways", Digital Equipment Corporation, March 1989.
 Postel, J., Editor, "Internet Protocol - DARPA Internet Program
Protocol Specification", STD 5, RFC 791, USC/Information Sciences
Institute, September 1981.
 Postel, J., Editor, "Transmission Control Protocol - DARPA
Internet Program Protocol Specification", STD 7, RFC 793,
USC/Information Sciences Institute, September 1981.
 Clark, D., "IP Datagram Reassembly Algorithms", RFC 815, MIT
Laboratory for Computer Science/Computer Systems and
Communications Group, July 1982.
G. Paul Ziemba
2115 O'Nel Drive
San Jose, CA 95131
1275A Malvern Rd
Melbourne, Vic 3144
cisco Systems, Inc.
170 W. Tasman Dr.
San Jose, CA 95028