idnits 2.17.1 draft-ietf-mptcp-threat-03.txt: Checking boilerplate required by RFC 5378 and the IETF Trust (see https://trustee.ietf.org/license-info): ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/1id-guidelines.txt: ---------------------------------------------------------------------------- No issues found here. Checking nits according to https://www.ietf.org/id-info/checklist : ---------------------------------------------------------------------------- ** The document seems to lack an IANA Considerations section. (See Section 2.2 of https://www.ietf.org/id-info/checklist for how to handle the case when there are no actions for IANA.) Miscellaneous warnings: ---------------------------------------------------------------------------- == The copyright year in the IETF Trust and authors Copyright Line does not match the current year -- The document date (October 4, 2010) is 4953 days in the past. Is this intentional? Checking references for intended status: Informational ---------------------------------------------------------------------------- == Unused Reference: 'RFC2960' is defined on line 695, but no explicit reference was found in the text -- Obsolete informational reference (is this intentional?): RFC 3775 (Obsoleted by RFC 6275) -- Obsolete informational reference (is this intentional?): RFC 2960 (Obsoleted by RFC 4960) Summary: 1 error (**), 0 flaws (~~), 2 warnings (==), 3 comments (--). Run idnits with the --verbose option for more detailed information about the items above. -------------------------------------------------------------------------------- 2 Network Working Group M. Bagnulo 3 Internet-Draft UC3M 4 Intended status: Informational October 4, 2010 5 Expires: April 7, 2011 7 Threat Analysis for Multi-addressed/Multi-path TCP 8 draft-ietf-mptcp-threat-03 10 Abstract 12 Multi-addresses/Multi-path TCP (MPTCP for short) describes the 13 extensions proposed for TCP so that each endpoint of a given TCP 14 connection can use multiple IP addresses to exchange data (instead of 15 a single IP address per endpoint as currently defined). Such 16 extensions enable the exchange of segments using different source- 17 destination address pairs, resulting in the capability of using 18 multiple paths in a significant number of scenarios. In particular, 19 some level of multihoming and mobility support can be achieved 20 through these extensions. However, the support for multiple IP 21 addresses per endpoint may have implications on the security of the 22 resulting MPTCP protocol. This note includes a threat analysis for 23 MPTCP. 25 Status of this Memo 27 This Internet-Draft is submitted in full conformance with the 28 provisions of BCP 78 and BCP 79. 30 Internet-Drafts are working documents of the Internet Engineering 31 Task Force (IETF). Note that other groups may also distribute 32 working documents as Internet-Drafts. The list of current Internet- 33 Drafts is at http://datatracker.ietf.org/drafts/current/. 35 Internet-Drafts are draft documents valid for a maximum of six months 36 and may be updated, replaced, or obsoleted by other documents at any 37 time. It is inappropriate to use Internet-Drafts as reference 38 material or to cite them other than as "work in progress." 40 This Internet-Draft will expire on April 7, 2011. 42 Copyright Notice 44 Copyright (c) 2010 IETF Trust and the persons identified as the 45 document authors. All rights reserved. 47 This document is subject to BCP 78 and the IETF Trust's Legal 48 Provisions Relating to IETF Documents 49 (http://trustee.ietf.org/license-info) in effect on the date of 50 publication of this document. Please review these documents 51 carefully, as they describe your rights and restrictions with respect 52 to this document. Code Components extracted from this document must 53 include Simplified BSD License text as described in Section 4.e of 54 the Trust Legal Provisions and are provided without warranty as 55 described in the Simplified BSD License. 57 Table of Contents 59 1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 3 60 2. Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 61 3. Related work . . . . . . . . . . . . . . . . . . . . . . . . . 4 62 4. Basic MPTCP. . . . . . . . . . . . . . . . . . . . . . . . . . 6 63 5. Flooding attacks . . . . . . . . . . . . . . . . . . . . . . . 7 64 6. Hijacking attacks . . . . . . . . . . . . . . . . . . . . . . 10 65 6.1. Hijacking attacks to the Basic MPTCP protocol . . . . . . 10 66 6.2. Time-shifted hijacking attacks . . . . . . . . . . . . . . 12 67 6.3. NAT considerations . . . . . . . . . . . . . . . . . . . . 14 68 7. Reccommendation . . . . . . . . . . . . . . . . . . . . . . . 14 69 8. Security Considerations . . . . . . . . . . . . . . . . . . . 15 70 9. Contributors . . . . . . . . . . . . . . . . . . . . . . . . . 15 71 10. Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . 15 72 11. Informative References . . . . . . . . . . . . . . . . . . . . 15 73 Author's Address . . . . . . . . . . . . . . . . . . . . . . . . . 16 75 1. Introduction 77 Multi-addresses/Multi-path TCP (MPTCP for short) describes the 78 extensions proposed for TCP so that each endpoint of a given TCP 79 connection can use multiple IP addresses to exchange data (instead of 80 a single IP address per endpoint as currently defined). Such 81 extensions enable the exchange of segments using different source- 82 destination address pairs, resulting in the capability of using 83 multiple paths in a significant number of scenarios. In particular, 84 some level of multihoming and mobility support can be achieved 85 through these extensions. However, the support for multiple IP 86 addresses per endpoint may have implications on the security of the 87 resulting MPTCP protocol. This note includes a threat analysis for 88 MPTCP. 90 2. Scope 92 There are multiple ways to achieve Multi-path TCP. Essentially what 93 is needed is for different segments of the communication to be 94 forwarded through different paths by enabling the sender to specify 95 some form of path selector. There are multiple options for such path 96 selector, including the usage of different next hops, using tunnels 97 to different egress points and so on. In this note, we will focus on 98 a particular approach, namely MPTCP approaches that rely on the usage 99 of multiple IP address per endpoint and that use different source- 100 destination address pairs as a mean to express different paths. So, 101 in the rest of this note, the MPTCP expression will refer to this 102 Multi-addressed flavour of Multi-path TCP. 104 Scope of the analysis 106 In this note we perform a threat analysis for MPTCP. Introducing the 107 support of multiple addresses per endpoint in a single TCP connection 108 may result in additional vulnerabilities. The scope of this note is 109 to identify and characterize these new vulnerabilities. So, the 110 scope of the analysis is limited to the additional vulnerabilities 111 resulting from the multi-address support compared to the current TCP 112 protocol (where each endpoint only has one address available for use 113 per connection). In other words, a full analysis of the complete set 114 of threats is explicitly out of the scope. The goal of this analysis 115 is to help the MPTCP protocol designers to create a MPTCP that is as 116 secure as the current TCP. It is a non goal of this analysis to help 117 in the design of MPTCP that is more secure than regular TCP. 119 In particular, we will focus on attackers that are not along the 120 path, at least not during the whole duration of the connection. In 121 the current single path TCP, on-path attacker can launch a 122 significant number of attacks, including eavesdropping, connection 123 hijacking Man in the Middle attacks and so on. However, it is not 124 possible for the off-path attackers to launch such attacks. There is 125 a middle ground in case the attacker is located along the path for a 126 short period of time to launch the attack and then moves away, but 127 the attack effects still apply. These are the so-called time-shifted 128 attacks. Since these are not possible in today's TCP, we will also 129 consider them as part of the analysis. So, summarizing, we will 130 consider both attacks launched by off-path attackers and time-shifted 131 attacks. Attacks launched by on-path attackers are out of scope, 132 since they also apply to current single-path TCP. 133 It should be noted, however, that some current on-path attacks may 134 become more difficult with multi-path TCP, since an attacker (on a 135 single path) will not have visibility of the complete data stream. 137 3. Related work 139 There is significant amount of previous work in terms of analysis of 140 protocols that support address agility. In this section we present 141 the most relevant ones and we relate them to the current MPTCP 142 effort. 144 Most of the problems related to address agility have been deeply 145 analyzed and understood in the context of Route Optimization support 146 in Mobile IPv6 (MIPv6 RO) [RFC3775]. [RFC4225] includes the rational 147 for the design of the security of MIPv6 RO. All the attacks 148 described in the aforementioned analysis apply here and are an 149 excellent basis for our own analysis. The main differences are: 150 In MIPv6 RO, the address binding affects all the communications 151 involving an address, while in the MPTCP case, a single connection 152 is at stake. In other words, if a binding between two address is 153 created at the IP layer, this binding can and will affect all the 154 connections that involve those addresses. However, in MPTCP, if 155 an additional address is added to an ongoing TCP connection, the 156 additional address will/can only affect the connection at hand and 157 not other connections even if the same address is being used for 158 those other connections. The result is that in MPTCP there is 159 much less at stake and the resulting vulnerabilities are less. On 160 the other hand, it is very important to keep the assumption valid 161 that the address bindings for a given connection do not affect 162 other connections. If reusing of binding or security information 163 is to be considered, this assumption could be no longer valid and 164 the full impact of the vulnerabilities must be assessed. 165 In MIPv6 RO, there is the assumption that the original path 166 through which the connection has been established is always 167 available and in case it is not, the communication will be lost. 168 In MPTCP, it is an explicit goal to provide communication 169 resilience when one of the address pairs is no longer usable, so 170 it is not possible to leverage on the original address pair to be 171 always working. 172 MIPv6 RO is of course designed for IPv6 and it is an explicit goal 173 of MPTCP to support both IPv6 and IPv4. Some MIPv6 RO security 174 solutions rely on the usage of some characteristics of IPv6 (such 175 as the usage of CGAs [RFC3972]), which will not be usable in the 176 context of MPTCP. 177 As opposed to MPTCP, MIPv6 RO does not have a connection state 178 information, including sequence numbers, port that could be 179 leveraged to provide security in some form. 181 In the Shim6 [RFC5533] design, similar issues related to address 182 agility were considered and a threat analysis was also performed 183 [RFC4218]. The analysis performed for Shim6 also largely applies to 184 the MPTCP context, the main difference being: 185 Similarly to the MPTCP case, the Shim6 protocol is a layer 3 186 protocol so all the communications involving the target address 187 are at stake, as opposed to the MPTCP case, where the impact can 188 be limited to a single TCP connection. 189 Similarly to MIPv6 RO, Shim6 only uses IPv6 addresses as 190 identifiers and leverages on some of their properties to provide 191 the security, such as relying on CGAs or HBAs [RFC5535], which is 192 not possible in the MPTCP case where IPv4 addresses must be 193 supported. 194 Similarly to MIPv6 RO, Shim6 does not have a connection state 195 information, including sequence numbers, port that could be 196 leveraged to provide security in some form. 198 SCTP [RFC2960]is a transport protocol that supports multiple 199 addresses per endpoint and as such, the security implications are 200 very close to the ones of MPTCP. A security analysis, identifying a 201 set of attacks and proposed solutions was performed in [RFC5062]. 202 The results of this analysis apply directly to the case of MPTCP. 203 However, the analysis was performed after the base SCTP protocol was 204 designed and the goal of the document was essentially to improve the 205 security of SCTP. As such, the document is very specific to the 206 actual SCTP specification and relies on the SCTP messages and 207 behaviour to characterize the issues. While some them can be 208 translated to the MPTCP case, some may be caused by specific 209 behaviour of SCTP as defined. 211 So, the conclusion is that while we do have a significant amount of 212 previous work that is closely related and we can and will use it as a 213 basis for this analysis, there are a set of characteristics that are 214 specific to MPTCP that grant the need for a specific analysis for 215 MPTCP. The goal of this analysis is to help MPTCP protocol designers 216 to include a set of security mechanisms that prevent the introduction 217 of new vulnerabilities to the Internet due to the adoption of MPTCP. 219 4. Basic MPTCP. 221 As we stated earlier, the goal of this document is to serve as input 222 for MPTCP protocol designers to properly take into account the 223 security issues. As such, the analysis cannot be performed for a 224 specific MPTCP specification, but must be a general analysis that 225 applies to the widest possible set of MPTCP designs. In order to do 226 that, we will characterize what are the fundamental features that any 227 MPTCP protocol must provide and attempt to perform the security 228 implications only assuming those. In some cases, we will have a 229 design choice that will significantly influence the security aspects 230 of the resulting protocol. In that case we will consider both 231 options and try to characterize both designs. 233 We assume that any MPTCP will behave in the case of a single address 234 per endpoint as TCP. This means that a MPTCP connection will be 235 established by using the TCP 3-way handshake and will use a single 236 address pair. 238 The addresses used for the establishment of the connection do have a 239 special role in the sense that this is the address used as identifier 240 by the upper layers. In particular, the address used as destination 241 address in the SYN packet is the address that the application is 242 using to identify the peer and has been obtained either through the 243 DNS (with or without DNSSEC validation) or passed by a referral or 244 manually introduced by the user. As such, the initiator does have a 245 certain amount of trust in the fact that it is establishing a 246 communication with that particular address. If due to MPTCP, packets 247 end up being delivered to an alternative address, the trust that the 248 initiator has placed on that address would be deceived. In any case, 249 the adoption of MPTCP necessitates a slight evolution of the 250 traditional TCP trust model, in that the initiator is additionally 251 trusting the peer to provide additional addresses which it will trust 252 to the same degree as the original pair. An application or 253 implementation that cannot trust the peer in this way should not make 254 use of multiple paths. 256 During the 3-way handshake, the sequence number will be synchronized 257 for both ends, as in regular TCP. We assume that a MPTCP connection 258 will use a single sequence number for the data, even if the data is 259 exchanged through different paths, as MPTCP provides an in-order 260 delivery service of bytes 262 Once the connection is established, the MPTCP extensions can be used 263 to add addresses for each of the endpoints. In order to do that each 264 end will need to send a control message containing the additional 265 address(es). In order to associate the additional address to an 266 ongoing connection, the connection needs to be identified. We assume 267 that the connection can be identified by the 4-tuple of source 268 address, source port, destination address, destination port used for 269 the establishment of the connection. So, at least, the control 270 message that will convey the additional address information can also 271 contain the 4-tuple in order to inform about what connection the 272 address belong to (if no other connection identifier is defined). 273 There are two different ways to convey address information: 274 o Explicit mode: the control message contain a list of addresses. 275 o Implicit mode: the address added is the one included in the source 276 address field of the IP header 278 These two modes have different security properties for some type of 279 attacks. The explicit mode seems to be the more vulnerable to abuse. 280 In particular, the implicit mode may benefit from forms of ingress 281 filtering security, which would reduce the possibility of an attacker 282 to add any arbitrary address to an ongoing connection. However, it 283 should be noted that ingress filtering deployment is far from 284 universal, and as such it is unwise to rely on it as a basis for the 285 protection of MPTCP. 286 In addition, further consideration about the interaction between 287 ingress filtering and implicit mode signaling is needed in the 288 case that we need to remove an address that is no longer available 289 from the MPTCP connection. In particular a host attached to a 290 network that performs ingress filtering and using implicit 291 signaling would not be able to remove an address that is no longer 292 available (either because of a failure or due to a mobility event) 293 from an ongoing MPTCP connection. 295 In addition, we will assume that MPTCP will use all the address pairs 296 that it has available for sending packets and that it will distribute 297 the load based on congestion among the different paths. 299 5. Flooding attacks 301 The first type of attacks that are introduced by address agility are 302 the so called flooding (or bombing) attacks. The setup for this 303 attack is depicted in the following figure: 305 +--------+ (step 1) +------+ 306 |Attacker| ------------------------- |Source| 307 | A |IPA IPS| S | 308 +--------+ /+------+ 309 / 310 (step 2) / 311 / 312 v IPT 313 +------+ 314 |Target| 315 | T | 316 +------+ 318 The scenario consists of an attacker A who has an IP address IPA. A 319 server that can generate a significant amount of traffic (such as a 320 streaming server), called source S and that has IP address IPS. In 321 addition, we have the target of the flooding attack, target T which 322 has an IP address IPT. 324 In the first step of this attack (depicted as step 1 in the figure), 325 the attacker A establishes a MPTCP connection with the source of the 326 traffic server S and starts downloading a significant amount of 327 traffic. The initial connection only involves one IP address per 328 endpoint, namely IPA and IPS. Once that the download is on course, 329 the second step of the attack (depicted as step 2 in the figure) is 330 that the attacker A adds IPT as one of the available addresses for 331 the communication. How the additional address is added depends on 332 the MPTCP address management mode. In explicit address management, 333 the attacker A only needs to send a signaling packet conveying 334 address IPT. In implicit mode, the attacker A would need to send a 335 packet with IPT as the source address. Depending on whether ingress 336 filtering is deployed and the location of the attacker, it may be 337 possible or not for the attacker to send such packet. At this stage, 338 the MPTCP connection still has a single address for the Source S i.e. 339 IPS but has two addresses for the Attacker A, namely IPA and IPT. 340 The attacker now attempts to get the Source S to send the traffic of 341 the ongoing download to the Target T IP address i.e. IPT. The 342 attacker can do that by pretending that the path between IPA and IPT 343 is congested but that the path between IPS and IPT is not. In order 344 to do that, it needs to send ACKs for the data that flows through the 345 path between IPS and IPT and do not send ACKs for the data that is 346 sent to IPA. The actual details of this will depend on how the data 347 sent through the different paths is ACKed. One possibility is that 348 ACKs for the data sent using a given a given address pair should come 349 in packets containing the same address pair. If so, the attacker 350 would need to send ACKs using packets containing IPT as the source 351 address to keep the attack flowing. This may be possible or not 352 depending on the deployment of ingress filtering and the location of 353 the attacker. The attacker would also need to guess the sequence 354 number of the data being sent to the Target. Once the attacker 355 manages to perform these actions the attack is on place and the 356 download will hit the target. It should be noted that in this type 357 of attacks, the Source S still thinks it is sending packets to the 358 Attacker A while in reality it is sending the packet to Target T. 360 Once that the traffic from the Source S start hitting the Target T, 361 the target will react. In particular, since the packets are likely 362 to belong to a non existent TCP connection, the Target T will issue 363 RST packets. It is relevant then to understand how MPTCP reacts to 364 incoming RST packets. It seems that the at least the MPTCP that 365 receives a RST packet should terminate the packet exchange 366 corresponding to the particular address pair (maybe not the complete 367 MPTCP connection, but at least it should not send more packets with 368 the address pair involved in the RST packet). However, if the 369 attacker, before redirecting the traffic has managed to increase the 370 window size considerably, the flight size could be enough to impose a 371 significant amount of traffic to the Target node. There is a subtle 372 operation that the attacker needs to achieve in order to launch a 373 significant attack. On the one hand it needs to grow the window 374 enough so that the flight size is big enough to cause enough effect 375 and on the other hand the attacker needs to be able to simulate 376 congestion on the IPA-IPS path so that traffic is actually redirected 377 to the alternative path without significantly reducing the window. 378 This will heavily depend on how the coupling of the windows between 379 the different paths works, in particular how the windows are 380 increased. Some designs of the congestion control window coupling 381 could render this attack ineffective. 383 Previous protocols, such as MIPv6 RO and SCTP, that have to deal with 384 this type of attacks have done so by adding a reachability check 385 before actually sending data to a new address. In other words, the 386 solution used in other protocols, would include the Source S to 387 explicitly asking the host sitting in the new address (in this case 388 the Target T sitting in IPT) whether it is willing to accept packets 389 from the MPTCP connection identified by the 4-tuple IPA, port A, IPS, 390 port S. Since this is not part of the established connection that 391 Target T has, T would not accept the request and Source S would not 392 use IPT to send packets for this MPTCP connection. Usually, the 393 request also includes a nonce that cannot be guessed by the attacker 394 A so that it cannot fake the reply to the request easily. In 395 particular, In the case of SCTP, it sends a message with a 64-bit 396 nonce (in a HEARTBEAT). 398 One possible approach to do this reachability test would be to 399 perform a 3-way handshake for each new address pair that is going to 400 be used in a MPTCP connection. While there are other reasons for 401 doing this (such as NAT traversal), such approach would also act as a 402 reachability test and would prevent the flooding attacks described in 403 this section. 405 6. Hijacking attacks 407 6.1. Hijacking attacks to the Basic MPTCP protocol 409 The hijacking attacks essentially use the MPTCP address agility to 410 allow an attacker to hijack a connection. This means that the victim 411 of a connection thinks that it is talking to a peer, while it is 412 actually exchanging packets with the attacker. In some sense it is 413 the dual of the flooding attacks (where the victim thinks it is 414 exchanging packets with the attacker but in reality is sending the 415 packets to the target). 417 The scenario for a hijacking attack is described in the next figure. 419 +------+ +------+ 420 | Node | ------------------------- | Node | 421 | 1 |IP1 IP2| 2 | 422 +------+ /+------+ 423 / 424 / 425 / 426 v IPA 427 +--------+ 428 |Attacker| 429 | A | 430 +--------+ 432 In this case, we have a MPTCP connection established between Node 1 433 and Node 2. The connection is using only one address per endpoint, 434 namely IP1 and IP2. The attacker then launches the hijacking attack 435 by adding IPA as an additional address for Node 1. In this case, 436 there is not much difference between explicit or implicit address 437 management, since in both cases the Attacker A could easily send a 438 control packet adding the address IPA, either as control data or as 439 the source address of the control packet. In order to be able to 440 hijack the connection, the attacker needs to know the 4-tuple that 441 identifies the connection, including the pair of addresses and the 442 pair of ports. It seems reasonable to assume that knowing the source 443 and destination IP addresses and the port of the server side is 444 fairly easy for the attacker. Learning the port of the client (i.e. 445 of the initiator of the connection) may prove to be more challenging. 446 The attacker would need to guess what the port is or to learn it by 447 intercepting the packets. Assuming that the attacker can gather the 448 4-tuple and issue the message adding IPA to the addresses available 449 for the MPTCP connection, then the attacker A has been able to 450 participate in the communication. In particular: 451 o Segments flowing from the Node 2:Depending how the usage of 452 addresses is defined, Node 2 will start using IPA to send data to. 453 In general, since the main goal is to achieve multi-path 454 capabilities, we can assume that unless there are already many IP 455 address pairs in use in the MPTCP connection, Node 2 will start 456 sending data to IPA. This means that part of the data of the 457 communication will reach the Attacker but probably not all of it. 458 This per se, already has negative effects, since Node 1 will not 459 receive all the data from Node 2. Moreover, from the application 460 perspective, this would result in DoS attack, since the byte flow 461 will stop waiting for the missing data. However, it is not enough 462 to achieve full hijacking of the connection, since part of data 463 will be still delivered to IP1, so it would reach Node 1 and not 464 the Attacker. In order for the attacker to receive all the data 465 of the MPTCP connection, the Attacker must somehow remove IP1 of 466 the set of available addresses for the connection. in the case of 467 implicit address management, this operation is likely to imply 468 sending a termination packet with IP1 as source address, which may 469 or not be possible for the attacker depending on whether ingress 470 filtering is in place and the location of the attacker. If 471 explicit address management is used, then the attacker will send a 472 remove address control packet containing IP1. The result is that 473 once IP1 is removed, all the data sent by Node 2 will reach the 474 Attacker and the incoming traffic has been hijacked. 475 o Segments flowing to the Node 2: As soon as IPA is accepted by Node 476 2 as part of the address set for the MPTCP connection, the 477 Attacker can send packets using IPA and those packets will be 478 considered by Node 2 as part of MPTCP connection. This means that 479 the attacker will be able to inject data into the MPTCP 480 connection, so from this perspective, the attacker has hijacked 481 part of the outgoing traffic. However, Node 1 would still be able 482 to send traffic that will be received by Node 2 as part of the 483 MPTCP connection. This means that there will be two source of 484 data i.e. Node 1 and the attacker, potentially preventing the 485 full hijacking of the outgoing traffic by the attacker. In order 486 to achieve a full hijacking, the attacker would need to remove IP1 487 from the set of available addresses. This can be done using the 488 same techniques described in the previous paragraph. 490 A related attack that can be achieved using similar techniques would 491 be a Man in the Middle (MitM) attack. The scenario for the attack is 492 depicted in the figure below. 494 +------+ +------+ 495 | Node | --------------- | Node | 496 | 1 |IP1 IP2| 2 | 497 +------+ \ /+------+ 498 \ / 499 \ / 500 \ / 501 v IPA v 502 +--------+ 503 |Attacker| 504 | A | 505 +--------+ 507 In this case, there is an established connection between Node 1 and 508 Node 2. The Attacker A will use the MPTCP address agility 509 capabilities to place itself as a MitM. In order to do so, it will 510 add IP address IPA as an additional address for the MPTCP connection 511 on both Node 1 and Node 2. This is essentially the same technique 512 described earlier in this section, only that it is used against both 513 nodes involved in the communication. The main difference is that in 514 this case, the attacker can simply sniff the content of the 515 communication that is forwarded through it and in turn forward the 516 data to the peer of the communication. The result is that the 517 attacker can place himself in the middle of the communication and 518 sniff part of the traffic unnoticed. Similar considerations about 519 how the attacker can manage to get to see all the traffic by removing 520 the genuine address of the peer apply. 522 6.2. Time-shifted hijacking attacks 524 A simple way to prevent off-path attackers to launch hijacking 525 attacks is to provide security of the control messages that add and 526 remove addresses by the usage of a cookie. In this type of 527 approaches, the peers involved in the MPTCP connection agree on a 528 cookie, that is exchanged in plain text during the establishment of 529 the connection and that needs to be presented in every control packet 530 that adds or removes an address for any of the peers. The result is 531 that the attacker needs to know the cookie in order to launch any of 532 the hijacking attacks described earlier. This implies that off path 533 attackers can no longer perform the hijacking attacks and that only 534 on-path attackers can do so, so one may consider that a cookie based 535 approach to secure MPTCP connection results in similar security than 536 current TCP. While it is close, it is not entirely true. 538 The main difference between the security of a MPTCP protocol secured 539 through cookies and the current TCP protocol are the time shifted 540 attacks. As we described earlier, a time shifted attack is one where 541 the attacker is along the path during a period of time, and then 542 moves away but the effects of the attack still remains, after the 543 attacker is long gone. In the case of a MPTCP protocol secured 544 through the usage of cookies, the attacker needs to be along the path 545 until the cookie is exchanged. After the attacker has learnt the 546 cookie, it can move away from the path and can still launch the 547 hijacking attacks described in the previous section. 549 There are several type of approaches that provide some protection 550 against hijacking attacks and that are vulnerable to some forms of 551 time-shifted attacks. We will next present some general taxonomy of 552 solutions and we describe the residual threats: 553 o Cookie-based solution: As we described earlier, one possible 554 approach is to use a cookie, that is sent in clear text in every 555 MPTCP control message that adds a new address to the existing 556 connection. The residual threat in this type of solution is that 557 any attacker that can sniff any of these control messages will 558 learn the cookie and will be able to add new addresses at any 559 given point in the lifetime of the connection. Moreover, the 560 endpoints will not detect the attack since the original cookie is 561 being used by the attacker. Summarizing, the vulnerability window 562 of this type of attacks includes all the flow establishment 563 exchanges and it is undetectable by the endpoints. 564 o Shared secret exchanged in plain text: An alternative option that 565 is somehow more secure than the cookie based approach is to 566 exchange a key in clear text during the establishment of the first 567 subflow and then validate the following subflows by using an keyed 568 HMAC signature using the shared key. This solution would be 569 vulnerable to attackers sniffing the message exchange for the 570 establishment of the first subflow, but after that, the shared key 571 is not transmitted any more, so the attacker cannot learn it 572 through sniffing any other message. Unfortunately, in order to be 573 compatible with NATs (see analysis below) even though this 574 approach includes a keyed HMAC signature, this signature cannot 575 cover the IP address that is being added. This basically means 576 that this type of approaches are also vulnerable to integrity 577 attacks of the exchanged messages. This means that even though 578 the attacker cannot learn the shared key by sniffing the 579 subsequent subflow establishment, the attacker can modify the 580 subflow establishment message and change the address that is being 581 added. So, the vulnerability window for confidentially to the 582 shared key is limited to the establishment of the first subflow, 583 but the vulnerability window for integrity attacks still includes 584 all the subflow establishment exchanges. These attacks are still 585 undetectable by the endpoints. It should be noted that the SCTP 586 security falls in this category. 587 o Strong crypto anchor exchange. another approach that could be used 588 would be to exchange some strong crypto anchor while the 589 establishment of the first subflow, such as a public key or a hash 590 chain anchor. In this case, subsequent subflows could be 591 protected by using the crypto material associated to that anchor. 592 An attacker in this case would need to change the crypto material 593 exchanged in the connection establishment phase. As a result the 594 vulnerability window for forging the crypto anchor is limited to 595 the initial connection establishment exchange. Similarly to the 596 previous case, due to NAT traversal considerations, the 597 vulnerability window for integrity attacks include all the subflow 598 establishment exchanges. As opposed to the previous one, because 599 the attacker needs to change the crypto anchor, this approach are 600 detectable by the endpoints, if they communicate directly. 602 6.3. NAT considerations 604 In order to be widely adopted MPTCP must work through NATs. NATs are 605 an interesting device from a security perspective. In terms of MPTCP 606 they essentially behave as a Man-in-the-middle attacker. As we have 607 described earlier, MPTCP security goal is to prevent from any 608 attacker to insert their addresses as valid addresses for a given 609 MPTCP connection. But that is exactly what a NAT does, they modify 610 the addresses. So, if MPTCP is to work through NATs, MPTCP must 611 accept address rewritten by NATs as valid addresses for a given 612 session. The most direct corollary is that the MPTCP messages that 613 add addresses in the implicit mode (i.e. the SYN of new subflows) 614 cannot be protected against integrity attacks, since they must allow 615 for NATs to change their addresses. This basically rules out any 616 solution that would rely on providing integrity protection to prevent 617 an attacker from changing the address used in a subflow establishment 618 exchange. This implies that alternative creative mechanisms are 619 needed to protect from integrity attacks to the MPTCP signaling that 620 adds new addresses to a connection. It is far from obvious how one 621 such creative approach could look like at this point. 623 7. Reccommendation 625 The presented analysis shows that there is a tradeoff between the 626 complexity of the security solution and the residual threats. In 627 order to define a proper security solution, we need to assess the 628 tradeoff and make a recommendation. After evaluating the different 629 aspects in the MPTCP WG, our conclusion are: 630 MPTCP should implement some form of reachabilty check using a 631 random nonce (e.g. TCP 3-way handshake) before adding a new 632 address to an ongoing communication in order to prevent flooding 633 attacks, 634 The default security mechanisms for MPTCP should be to exchange a 635 key in the establishment of the first subflow and then secure 636 following address additions by using a keyed HMAC using the 637 exchanged key. 638 MPTCP security mechanism should support using a pre-shared key 639 to be used in the keyed HMAC, providing a higher level of 640 protection than the previous one. 641 A mechanism to prevent replay attacks using these messages 642 should be provided e.g. a sequence number protected by the HMAC 643 The MPTCP protocol should be extensible and it should able to 644 accommodate multiple security solutions, in order to enable the 645 usage of more secure mechanisms if needed. 647 8. Security Considerations 649 This note contains a security analysis for MPTCP, so no further 650 security considerations need to be described in this section 652 9. Contributors 654 Alan Ford - Roke Manor Research Ltd. 656 10. Acknowledgments 658 Rolf Winter, Randall Stewart, Andrew McDonald, Michael Tuexen 659 reviewed an earlier version of this document and provided comments to 660 improve it. 662 Mark Handley pointed out the problem with NATs and integrity 663 protection of MPTCP signaling. 665 Marcelo Bagnulo is partly funded by Trilogy, a research project 666 supported by the European Commission under its Seventh Framework 667 Program. 669 11. Informative References 671 [RFC4225] Nikander, P., Arkko, J., Aura, T., Montenegro, G., and E. 672 Nordmark, "Mobile IP Version 6 Route Optimization Security 673 Design Background", RFC 4225, December 2005. 675 [RFC4218] Nordmark, E. and T. Li, "Threats Relating to IPv6 676 Multihoming Solutions", RFC 4218, October 2005. 678 [RFC3972] Aura, T., "Cryptographically Generated Addresses (CGA)", 679 RFC 3972, March 2005. 681 [RFC5062] Stewart, R., Tuexen, M., and G. Camarillo, "Security 682 Attacks Found Against the Stream Control Transmission 683 Protocol (SCTP) and Current Countermeasures", RFC 5062, 684 September 2007. 686 [RFC5535] Bagnulo, M., "Hash-Based Addresses (HBA)", RFC 5535, 687 June 2009. 689 [RFC3775] Johnson, D., Perkins, C., and J. Arkko, "Mobility Support 690 in IPv6", RFC 3775, June 2004. 692 [RFC5533] Nordmark, E. and M. Bagnulo, "Shim6: Level 3 Multihoming 693 Shim Protocol for IPv6", RFC 5533, June 2009. 695 [RFC2960] Stewart, R., Xie, Q., Morneault, K., Sharp, C., 696 Schwarzbauer, H., Taylor, T., Rytina, I., Kalla, M., 697 Zhang, L., and V. Paxson, "Stream Control Transmission 698 Protocol", RFC 2960, October 2000. 700 Author's Address 702 Marcelo Bagnulo 703 Universidad Carlos III de Madrid 704 Av. Universidad 30 705 Leganes, Madrid 28911 706 SPAIN 708 Phone: 34 91 6248814 709 Email: marcelo@it.uc3m.es 710 URI: http://www.it.uc3m.es