Network Working Group R. Sparks, Ed. Request for Comments: 5393 Tekelec Updates: 3261 S. Lawrence Category: Standards Track Nortel Networks, Inc. A. Hawrylyshen Ditech Networks Inc. B. Campen Tekelec December 2008 Addressing an Amplification Vulnerability in Session Initiation Protocol (SIP) Forking Proxies Status of This Memo This document specifies an Internet standards track protocol for the Internet community, and requests discussion and suggestions for improvements. Please refer to the current edition of the "Internet Official Protocol Standards" (STD 1) for the standardization state and status of this protocol. Distribution of this memo is unlimited. Copyright Notice Copyright (c) 2008 IETF Trust and the persons identified as the document authors. All rights reserved. This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (http://trustee.ietf.org/ license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document. Abstract This document normatively updates RFC 3261, the Session Initiation Protocol (SIP), to address a security vulnerability identified in SIP proxy behavior. This vulnerability enables an attack against SIP networks where a small number of legitimate, even authorized, SIP requests can stimulate massive amounts of proxy-to-proxy traffic. This document strengthens loop-detection requirements on SIP proxies when they fork requests (that is, forward a request to more than one destination). It also corrects and clarifies the description of the loop-detection algorithm such proxies are required to implement. Additionally, this document defines a Max-Breadth mechanism for limiting the number of concurrent branches pursued for any given request. Sparks, et al. Standards Track [Page 1] RFC 5393 Amplification Vulnerability in SIP December 2008 Table of Contents 1. Introduction ....................................................3 2. Conventions and Definitions .....................................3 3. Vulnerability: Leveraging Forking to Flood a Network ............3 4. Updates to RFC 3261 .............................................7 4.1. Strengthening the Requirement to Perform Loop Detection ....7 4.2. Correcting and Clarifying the RFC 3261 Loop-Detection Algorithm ...................................7 4.2.1. Update to Section 16.6 ..............................7 4.2.2. Update to Section 16.3 ..............................8 4.2.3. Impact of Loop Detection on Overall Network Performance .........................................9 4.2.4. Note to Implementers ................................9 5. Max-Breadth ....................................................10 5.1. Overview ..................................................10 5.2. Examples ..................................................11 5.3. Formal Mechanism ..........................................12 5.3.1. Max-Breadth Header Field ...........................12 5.3.2. Terminology ........................................13 5.3.3. Proxy Behavior .....................................13 5.3.3.1. Reusing Max-Breadth .......................14 5.3.4. UAC Behavior .......................................14 5.3.5. UAS Behavior .......................................14 5.4. Implementer Notes .........................................14 5.4.1. Treatment of CANCEL ................................14 5.4.2. Reclamation of Max-Breadth on 2xx Responses ........14 5.4.3. Max-Breadth and Automaton UAs ......................14 5.5. Parallel and Sequential Forking ...........................15 5.6. Max-Breadth Split Weight Selection ........................15 5.7. Max-Breadth's Effect on Forking-Based Amplification Attacks .....................................15 5.8. Max-Breadth Header Field ABNF Definition ..................16 6. IANA Considerations ............................................16 6.1. Max-Breadth Header Field ..................................16 6.2. 440 Max-Breadth Exceeded Response .........................16 7. Security Considerations ........................................16 7.1. Alternate Solutions That Were Considered and Rejected .....17 8. Acknowledgments ................................................19 9. References .....................................................19 9.1. Normative References ......................................19 9.2. Informative References ....................................19 Sparks, et al. Standards Track [Page 2] RFC 5393 Amplification Vulnerability in SIP December 2008 1. Introduction Interoperability testing uncovered a vulnerability in the behavior of forking SIP proxies as defined in [RFC3261]. This vulnerability can be leveraged to cause a small number of valid SIP requests to generate an extremely large number of proxy-to-proxy messages. A version of this attack demonstrates fewer than ten messages stimulating potentially 2^71 messages. This document specifies normative changes to the SIP protocol to address this vulnerability. According to this update, when a SIP proxy forks a request to more than one destination, it is required to ensure it is not participating in a request loop. This normative update alone is insufficient to protect against crafted variations of the attack described here involving multiple Addresses of Record (AORs). To further address the vulnerability, this document defines the Max-Breadth mechanism to limit the total number of concurrent branches caused by a forked SIP request. The mechanism only limits concurrency. It does not limit the total number of branches a request can traverse over its lifetime. The mechanisms in this update will protect against variations of the attack described here that use a small number of resources, including most unintentional self-inflicted variations that occur through accidental misconfiguration. However, an attacker with access to a sufficient number of distinct resources will still be able to stimulate a very large number of messages. The number of concurrent messages will be limited by the Max-Breadth mechanism, so the entire set will be spread out over a long period of time, giving operators better opportunity to detect the attack and take corrective measures outside the protocol. Future protocol work is needed to prevent this form of the attack. 2. Conventions and Definitions The key words "MUST", "MUST NOT", "REQUIRED", "SHALL", "SHALL NOT", "SHOULD", "SHOULD NOT", "RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be interpreted as described in RFC 2119 [RFC2119]. 3. Vulnerability: Leveraging Forking to Flood a Network This section describes setting up an attack with a simplifying assumption: that two accounts on each of two different RFC 3261 compliant proxy/registrar servers that do not perform loop detection are available to an attacker. This assumption is not necessary for the attack but makes representing the scenario simpler. The same attack can be realized with a single account on a single server. Sparks, et al. Standards Track [Page 3] RFC 5393 Amplification Vulnerability in SIP December 2008 Consider two proxy/registrar services, P1 and P2, and four Addresses of Record, a@P1, b@P1, a@P2, and b@P2. Using normal REGISTER requests, establish bindings to these AORs as follows (non-essential details elided): REGISTER sip:P1 SIP/2.0 To: Contact: , REGISTER sip:P1 SIP/2.0 To: Contact: , REGISTER sip:P2 SIP/2.0 To: Contact: , REGISTER sip:P2 SIP/2.0 To: Contact: , With these bindings in place, introduce an INVITE request to any of the four AORs, say a@P1. This request will fork to two requests handled by P2, which will fork to four requests handled by P1, which will fork to eight messages handled by P2, and so on. This message flow is represented in Figure 1. | a@P1 / \ / \ / \ / \ a@P2 b@P2 / \ / \ / \ / \ / \ / \ a@P1 b@P1 a@P1 b@P1 / \ / \ / \ / \ a@P2 b@P2 a@P2 b@P2 a@P2 b@P2 a@P2 b@P2 /\ /\ /\ /\ /\ /\ /\ /\ . . . Figure 1: Attack Request Propagation Sparks, et al. Standards Track [Page 4] RFC 5393 Amplification Vulnerability in SIP December 2008 Requests will continue to propagate down this tree until Max-Forwards reaches zero. If the endpoint and two proxies involved follow RFC 3261 recommendations, the tree will be 70 rows deep, representing 2^71-1 requests. The actual number of messages may be much larger if the time to process the entire tree's worth of requests is longer than Timer C at either proxy. In this case, a storm of 408 responses and/or a storm of CANCEL requests will also be propagating through the tree along with the INVITE requests. Remember that there are only two proxies involved in this scenario - each having to hold the state for all the transactions it sees (at least 2^70 simultaneously active transactions near the end of the scenario). The attack can be simplified to one account at one server if the service can be convinced that contacts with varying attributes (parameters, schemes, embedded headers) are sufficiently distinct, and these parameters are not used as part of AOR comparisons when forwarding a new request. Since RFC 3261 mandates that all URI parameters must be removed from a URI before looking it up in a location service and that the URIs from the Contact header field are compared using URI equality, the following registration should be sufficient to set up this attack using a single REGISTER request to a single account: REGISTER sip:P1 SIP/2.0 To: Contact: , This attack was realized in practice during one of the SIP Interoperability Test (SIPit) sessions. The scenario was extended to include more than two proxies, and the participating proxies all limited Max-Forwards to be no larger than 20. After a handful of messages to construct the attack, the participating proxies began bombarding each other. Extrapolating from the several hours the experiment was allowed to run, the scenario would have completed in just under 10 days. Had the proxies used the RFC 3261 recommended Max-Forwards value of 70, and assuming they performed linearly as the state they held increased, it would have taken 3 trillion years to complete the processing of the single INVITE request that initiated the attack. It is interesting to note that a few proxies rebooted during the scenario and rejoined in the attack when they restarted (as long as they maintained registration state across reboots). This points out that if this attack were launched on the Internet at large, it might require coordination among all the affected elements to stop it. Loop detection, as specified in this document, at any of the proxies in the scenarios described so far would have stopped the attack immediately. (If all the proxies involved implemented this loop Sparks, et al. Standards Track [Page 5] RFC 5393 Amplification Vulnerability in SIP December 2008 detection, the total number of stimulated messages in the first scenario described would be reduced to 14; in the variation involving one server, the number of stimulated messages would be reduced to 10.) However, there is a variant of the attack that uses multiple AORs where loop detection alone is insufficient protection. In this variation, each participating AOR forks to all the other participating AORs. For small numbers of participating AORs (10, for example), paths through the resulting tree will not loop until very large numbers of messages have been generated. Acquiring a sufficient number of AORs to launch such an attack on networks currently available is quite feasible. In this scenario, requests will often take many hops to complete a loop, and there are a very large number of different loops that will occur during the attack. In fact, if N is the number of participating AORs, and provided N is less than or equal to Max- Forwards, the amount of traffic generated by the attack is greater than N!, even if all proxies involved are performing loop detection. Suppose we have a set of N AORs, all of which are set up to fork to the entire set. For clarity, assume AOR 1 is where the attack begins. Every permutation of the remaining N-1 AORs will play out, defining (N-1)! distinct paths, without repeating any AOR. Then, each of these paths will fork N ways one last time, and a loop will be detected on each of these branches. These final branches alone total N! requests ((N-1)! paths, with N forks at the end of each path). ___N____Requests_ | 1 | 1 | | 2 | 4 | | 3 | 15 | | 4 | 64 | | 5 | 325 | | 6 | 1956 | | 7 | 13699 | | 8 | 109600 | | 9 | 986409 | | 10 | 9864100 | Forwarded Requests vs. Number of Participating AORs In a network where all proxies are performing loop detection, an attacker is still afforded rapidly increasing returns on the number of AORs they are able to leverage. The Max-Breadth mechanism defined in this document is designed to limit the effectiveness of this variation of the attack. Sparks, et al. Standards Track [Page 6] RFC 5393 Amplification Vulnerability in SIP December 2008 In all of the scenarios, it is important to notice that at each forking proxy, an additional branch could be added pointing to a single victim (that might not even be a SIP-aware element), resulting in a massive amount of traffic being directed towards the victim from potentially as many sources as there are AORs participating in the attack. 4. Updates to RFC 3261 4.1. Strengthening the Requirement to Perform Loop Detection The following requirements mitigate the risk of a proxy falling victim to the attack described in this document. When a SIP proxy forks a particular request to more than one location, it MUST ensure that request is not looping through this proxy. It is RECOMMENDED that proxies meet this requirement by performing the loop-detection steps defined in this document. The requirement to use this document's refinement of the loop- detection algorithm from RFC 3261 is set at should-strength to allow for future Standards-Track mechanisms that will allow a proxy to determine it is not looping. For example, a proxy forking to destinations established using the sip-outbound mechanism [OUTBOUND] would know those branches will not loop. A SIP proxy forwarding a request to only one location MAY perform loop detection but is not required to. When forwarding to only one location, the amplification risk being exploited is not present, and the Max-Forwards mechanism will protect the network to the extent it was designed (always keep in mind the constant multiplier due to exhausting Max-Forwards while not forking). A proxy is not required to perform loop detection when forwarding a request to a single location even if it happened to have previously forked that request (and performed loop detection) in its progression through the network. 4.2. Correcting and Clarifying the RFC 3261 Loop-Detection Algorithm 4.2.1. Update to Section 16.6 This section replaces all of item 8 in Section 16.6 of RFC 3261 (item 8 begins on page 105 and ends on page 106 of RFC 3261). Sparks, et al. Standards Track [Page 7] RFC 5393 Amplification Vulnerability in SIP December 2008 8. Add a Via Header Field Value The proxy MUST insert a Via header field value into the copy before the existing Via header field values. The construction of this value follows the same guidelines of Section 8.1.1.7. This implies that the proxy will compute its own branch parameter, which will be globally unique for that branch, and will contain the requisite magic cookie. Note that following only the guidelines in Section 8.1.1.7 will result in a branch parameter that will be different for different instances of a spiraled or looped request through a proxy. Proxies required to perform loop detection by RFC 5393 have an additional constraint on the value they place in the Via header field. Such proxies SHOULD create a branch value separable into two parts in any implementation-dependent way. The remainder of this section's description assumes the existence of these two parts. If a proxy chooses to employ some other mechanism, it is the implementer's responsibility to verify that the detection properties defined by the requirements placed on these two parts are achieved. The first part of the branch value MUST satisfy the constraints of Section 8.1.1.7. The second part is used to perform loop detection and distinguish loops from spirals. This second part MUST vary with any field used by the location service logic in determining where to retarget or forward this request. This is necessary to distinguish looped requests from spirals by allowing the proxy to recognize if none of the values affecting the processing of the request have changed. Hence, the second part MUST depend at least on the received Request-URI and any Route header field values used when processing the received request. Implementers need to take care to include all fields used by the location service logic in that particular implementation. This second part MUST NOT vary with the request method. CANCEL and non-200 ACK requests MUST have the same branch parameter value as the corresponding request they cancel or acknowledge. This branch parameter value is used in correlating those requests at the server handling them (see Sections 17.2.3 and 9.2). 4.2.2. Update to Section 16.3 This section replaces all of item 4 in Section 16.3 of RFC 3261 (item 4 appears on page 95 of RFC 3261). Sparks, et al. Standards Track [Page 8] RFC 5393 Amplification Vulnerability in SIP December 2008 4. Loop-Detection Check Proxies required to perform loop detection by RFC 5393 MUST perform the following loop-detection test before forwarding a request. Each Via header field value in the request whose sent-by value matches a value placed into previous requests by this proxy MUST be inspected for the "second part" defined in Section 4.2.1 of RFC 5393. This second part will not be present if the message was not forked when that Via header field value was added. If the second field is present, the proxy MUST perform the second-part calculation described in Section 4.2.1 of RFC 5393 on this request and compare the result to the value from the Via header field. If these values are equal, the request has looped and the proxy MUST reject the request with a 482 (Loop Detected) response. If the values differ, the request is spiraling and processing continues to the next step. 4.2.3. Impact of Loop Detection on Overall Network Performance These requirements and the recommendation to use the loop-detection mechanisms in this document make the favorable trade of exponential message growth for work that is, at worst, order n^2 as a message crosses n proxies. Specifically, this work is order m*n where m is the number of proxies in the path that fork the request to more than one location. In practice, m is expected to be small. The loop-detection algorithm expressed in this document requires a proxy to inspect each Via element in a received request. In the worst case, where a message crosses N proxies, each of which loop detect, proxy k does k inspections, and the overall number of inspections spread across the proxies handling this request is the sum of k from k=1 to k=N which is N(N+1)/2. 4.2.4. Note to Implementers A common way to create the second part of the branch parameter value when forking a request is to compute a hash over the concatenation of the Request-URI, any Route header field values used during processing the request, and any other values used by the location service logic while processing this request. The hash should be chosen so that there is a low probability that two distinct sets of these parameters will collide. Because the maximum number of inputs that need to be compared is 70, the chance of a collision is low even with a relatively small hash value, such as 32 bits. CRC-32c as specified in [RFC4960] is a specific acceptable function, as is MD5 [RFC1321]. Note that MD5 is being chosen purely for non-cryptographic properties. An attacker who can control the inputs in order to produce a hash collision can attack the connection in a variety of other ways. When forming the second part using a hash, Sparks, et al. Standards Track [Page 9] RFC 5393 Amplification Vulnerability in SIP December 2008 implementations SHOULD include at least one field in the input to the hash that varies between different transactions attempting to reach the same destination to avoid repeated failure should the hash collide. The Call-ID and CSeq fields would be good inputs for this purpose. A common point of failure to interoperate at SIPit events has been due to parsers objecting to the contents of another element's Via header field values when inspecting the Via stack for loops. Implementers need to take care to avoid making assumptions about the format of another element's Via header field value beyond the basic constraints placed on that format by RFC 3261. In particular, parsing a header field value with unknown parameter names, parameters with no values, or parameter values with or without quoted strings must not cause an implementation to fail. Removing, obfuscating, or in any other way modifying the branch parameter values in Via header fields in a received request before forwarding it removes the ability for the node that placed that branch parameter into the message to perform loop detection. If two elements in a loop modify branch parameters this way, a loop can never be detected. 5. Max-Breadth 5.1. Overview The Max-Breadth mechanism defined here limits the total number of concurrent branches caused by a forked SIP request. With this mechanism, all proxyable requests are assigned a positive integral Max-Breadth value, which denotes the maximum number of concurrent branches this request may spawn through parallel forking as it is forwarded from its current point. When a proxy forwards a request, its Max-Breadth value is divided among the outgoing requests. In turn, each of the forwarded requests has a limit on how many concurrent branches it may spawn. As branches complete, their portion of the Max-Breadth value becomes available for subsequent branches, if needed. If there is insufficient Max-Breadth to carry out a desired parallel fork, a proxy can return the 440 (Max-Breadth Exceeded) response defined in this document. This mechanism operates independently from Max-Forwards. Max- Forwards limits the depth of the tree a request may traverse as it is forwarded from its origination point to each destination it is forked to. As Section 3 shows, the number of branches in a tree of even limited depth can be made large (exponential with depth) by leveraging forking. Each such branch has a pair of SIP transaction Sparks, et al. Standards Track [Page 10] RFC 5393 Amplification Vulnerability in SIP December 2008 state machines associated with it. The Max-Breadth mechanism limits the number of branches that are active (those that have running transaction state machines) at any given point in time. Max-Breadth does not prevent forking. It only limits the number of concurrent parallel forked branches. In particular, a Max-Breadth of 1 restricts a request to pure serial forking rather than restricting it from being forked at all. A client receiving a 440 (Max-Breadth Exceeded) response can infer that its request did not reach all possible destinations. Recovery options are similar to those when receiving a 483 (Too Many Hops) response, and include affecting the routing decisions through whatever mechanisms are appropriate to result in a less broad search, or refining the request itself before submission to make the search space smaller. 5.2. Examples UAC Proxy A Proxy B Proxy C | INVITE | | | | Max-Breadth: 60 | INVITE | | | Max-Forwards: 70 | Max-Breadth: 30 | | |-------------------->| Max-Forwards: 69 | | | |------------------->| | | | INVITE | | | | Max-Breadth: 30 | | | | Max-Forwards: 69 | | | |--------------------------------------->| | | | | Parallel Forking UAC Proxy A Proxy B Proxy C | INVITE | | | | Max-Breadth: 60 | INVITE | | | Max-Forwards: 70 | Max-Breadth: 60 | | |-------------------->| Max-Forwards: 69 | | | |------------------->| | | | some error response| | | |<-------------------| | | | INVITE | | | | Max-Breadth: 60 | | | | Max-Forwards: 69 | | | |--------------------------------------->| | | | | Sequential Forking Sparks, et al. Standards Track [Page 11] RFC 5393 Amplification Vulnerability in SIP December 2008 UAC Proxy A Proxy B Proxy C | INVITE | | | | Max-Breadth: 60 | INVITE | | | Max-Forwards: 70 | Max-Breadth: 60 | INVITE | |-------------------->| Max-Forwards: 69 | Max-Breadth: 60 | | |------------------->| Max-Forwards: 68 | | | |------------------>| | | | | | | | | | | | | No Forking MB == Max-Breadth MF == Max-Forwards | MB: 4 | MF: 5 MB: 2 P MB: 2 MF: 4 / \ MF: 4 +---------------+ +------------------+ MB: 1 P MB: 1 MB: 1 P MB: 1 MF: 3 / \ MF: 3 MF: 3 / \ MF: 3 +---+ +-------+ +----+ +-------+ P P P P MB: 1 | MB: 1 | MB: 1 | MB: 1 | MF: 2 | MF: 2 | MF: 2 | MF: 2 | P P P P MB: 1 | MB: 1 | MB: 1 | MB: 1 | MF: 1 | MF: 1 | MF: 1 | MF: 1 | P P P P . . . Max-Breadth and Max-Forwards Working Together 5.3. Formal Mechanism 5.3.1. Max-Breadth Header Field The Max-Breadth header field takes a single positive integer as its value. The Max-Breadth header field value takes no parameters. Sparks, et al. Standards Track [Page 12] RFC 5393 Amplification Vulnerability in SIP December 2008 5.3.2. Terminology For each "response context" (see Section 16 of [RFC3261]) in a proxy, this mechanism defines two positive integral values: Incoming Max- Breadth and Outgoing Max-Breadth. Incoming Max-Breadth is the value in the Max-Breadth header field in the request that formed the response context. Outgoing Max-Breadth is the sum of the Max-Breadth header field values in all forwarded requests in the response context that have not received a final response. 5.3.3. Proxy Behavior If a SIP proxy receives a request with no Max-Breadth header field value, it MUST add one, with a value that is RECOMMENDED to be 60. Proxies MUST have a maximum allowable Incoming Max-Breadth value, which is RECOMMENDED to be 60. If this maximum is exceeded in a received request, the proxy MUST overwrite it with a value that SHOULD be no greater than its allowable maximum. All proxied requests MUST contain a single Max-Breadth header field value. SIP proxies MUST NOT allow the Outgoing Max-Breadth to exceed the Incoming Max-Breadth in a given response context. If a SIP proxy determines a response context has insufficient Incoming Max-Breadth to carry out a desired parallel fork, and the proxy is unwilling/unable to compensate by forking serially or sending a redirect, that proxy MUST return a 440 (Max-Breadth Exceeded) response. Notice that these requirements mean a proxy receiving a request with a Max-Breadth of 1 can only fork serially, but it is not required to fork at all -- it can return a 440 instead. Thus, this mechanism is not a tool a user agent can use to force all proxies in the path of a request to fork serially. A SIP proxy MAY distribute Max-Breadth in an arbitrary fashion between active branches. A proxy SHOULD NOT use a smaller amount of Max-Breadth than was present in the original request unless the Incoming Max-Breadth exceeded the proxy's maximum acceptable value. A proxy MUST NOT decrement Max-Breadth for each hop or otherwise use it to restrict the "depth" of a request's propagation. Sparks, et al. Standards Track [Page 13] RFC 5393 Amplification Vulnerability in SIP December 2008 5.3.3.1. Reusing Max-Breadth Because forwarded requests that have received a final response do not count towards the Outgoing Max-Breadth, whenever a final response arrives, the Max-Breadth that was used on that branch becomes available for reuse. Proxies SHOULD be prepared to reuse this Max- Breadth in cases where there may be elements left in the target-set. 5.3.4. UAC Behavior A User Agent Client (UAC) MAY place a Max-Breadth header field value in outgoing requests. If so, this value is RECOMMENDED to be 60. 5.3.5. UAS Behavior This mechanism does not affect User Agent Server (UAS) behavior. A UAS receiving a request with a Max-Breadth header field will ignore that field while processing the request. 5.4. Implementer Notes 5.4.1. Treatment of CANCEL Since CANCEL requests are never proxied, a Max-Breadth header field value is meaningless in a CANCEL request. Sending a CANCEL in no way affects the Outgoing Max-Breadth in the associated INVITE response context. Receiving a CANCEL in no way affects the Incoming Max- Breadth of the associated INVITE response context. 5.4.2. Reclamation of Max-Breadth on 2xx Responses Whether 2xx responses free up Max-Breadth is mostly a moot issue, since proxies are forbidden to start new branches in this case. But, there is one caveat. A proxy may receive multiple 2xx responses for a single forwarded INVITE request. Also, [RFC2543] implementations may send back a 6xx followed by a 2xx on the same branch. Implementations that subtract from the Outgoing Max-Breadth when they receive a 2xx response to an INVITE request must be careful to avoid bugs caused by subtracting multiple times for a single branch. 5.4.3. Max-Breadth and Automaton UAs Designers of automaton user agents (UAs) (including B2BUAs, gateways, exploders, and any other element that programmatically sends requests as a result of incoming SIP traffic) should consider whether Max- Breadth limitations should be placed on outgoing requests. For example, it is reasonable to design B2BUAs to carry the Max-Breadth value from incoming requests into requests that are sent as a result. Sparks, et al. Standards Track [Page 14] RFC 5393 Amplification Vulnerability in SIP December 2008 Also, it is reasonable to place Max-Breadth constraints on sets of requests sent by exploders when they may be leveraged in an amplification attack. 5.5. Parallel and Sequential Forking Inherent in the definition of this mechanism is the ability of a proxy to reclaim apportioned Max-Breadth while forking sequentially. The limitation on outgoing Max-Breadth is applied to concurrent branches only. For example, if a proxy receives a request with a Max-Breadth of 4 and has 8 targets to forward it to, that proxy may parallel fork to 4 of these targets initially (each with a Max-Breadth of 1, totaling an Outgoing Max-Breadth of 4). If one of these transactions completes with a failure response, the outgoing Max-Breadth drops to 3, allowing the proxy to forward to one of the 4 remaining targets (again, with a Max-Breadth of 1). 5.6. Max-Breadth Split Weight Selection There are a variety of mechanisms for controlling the weight of each fork branch. Fork branches that are given more Max-Breadth are more likely to complete quickly (because it is less likely that a proxy down the line will be forced to fork sequentially). By the same token, if it is known that a given branch will not fork later on, a Max-Breadth of 1 may be assigned with no ill effect. This would be appropriate, for example, if a proxy knows the branch is using the SIP outbound extension [OUTBOUND]. 5.7. Max-Breadth's Effect on Forking-Based Amplification Attacks Max-Breadth limits the total number of active branches spawned by a given request at any one time, while placing no constraint on the distance (measured in hops) that the request can propagate. (i.e., receiving a request with a Max-Breadth of 1 means that any forking must be sequential, not that forking is forbidden) This limits the effectiveness of any amplification attack that leverages forking because the amount of state/bandwidth needed to process the traffic at any given point in time is capped. Sparks, et al. Standards Track [Page 15] RFC 5393 Amplification Vulnerability in SIP December 2008 5.8. Max-Breadth Header Field ABNF Definition This specification extends the grammar for the Session Initiation Protocol by adding an extension-header. The ABNF [RFC5234] definition is as follows. Max-Breadth = "Max-Breadth" HCOLON 1*DIGIT 6. IANA Considerations This specification registers a new SIP header field and a new SIP response according to the processes defined in [RFC3261]. 6.1. Max-Breadth Header Field This information appears in the Header Fields sub-registry of the SIP Parameters registry. RFC 5393 (this specification) Header Field Name: Max-Breadth Compact Form: none 6.2. 440 Max-Breadth Exceeded Response This information appears in the Response Codes sub-registry of the SIP Parameters registry. Response code: 440 Default Reason Phrase: Max-Breadth Exceeded 7. Security Considerations This document is entirely about documenting and addressing a vulnerability in SIP proxies as defined by RFC 3261 that can lead to an exponentially growing message exchange attack. The Max-Breadth mechanism defined here does not decrease the aggregate traffic caused by the forking-loop attack. It only serves to spread the traffic caused by the attack over a longer period by limiting the number of concurrent branches that are being processed at the same time. An attacker could pump multiple requests into a network that uses the Max-Breadth mechanism and gradually build traffic to unreasonable levels. Deployments should monitor carefully and react to gradual increases in the number of concurrent outstanding transactions related to a given resource to protect Sparks, et al. Standards Track [Page 16] RFC 5393 Amplification Vulnerability in SIP December 2008 against this possibility. Operators should anticipate being able to temporarily disable any resources identified as being used in such an attack. A rapid increase in outstanding concurrent transactions system-wide may be an indication of the presence of this kind of attack across many resources. Deployments in which it is feasible for an attacker to obtain a very large number of resources are particularly at risk. If detecting and intervening in each instance of the attack is insufficient to reduce the load, overload may occur. Implementers and operators are encouraged to follow the recommendations being developed for handling overload conditions (see [REQS] and [DESIGN]). Designers of protocol gateways should consider the implications of this kind of attack carefully. As an example, if a message transits from a SIP network into the Public Switched Telephone Network (PSTN) and subsequently back into a SIP network, and information about the history of the request on either side of the protocol translation is lost, it becomes possible to construct loops that neither Max- Forwards nor loop detection can protect against. This, combined with forking amplification on the SIP side of the loop, will result in an attack as described in this document that the mechanisms here will not abate, not even to the point of limiting the number of concurrent messages in the attack. These considerations are particularly important for designers of gateways from SIP to SIP (as found in B2BUAs, for example). Many existing B2BUA implementations are under some pressure to hide as much information about the two sides communicating with them as possible. Implementers of such implementations may be tempted to remove the data that might be used by the loop-detection, Max-Forwards, or Max-Breadth mechanisms at other points in the network, taking on the responsibility for detecting loops (or forms of this attack). However, if two such implementations are involved in the attack, neither will be able to detect it. 7.1. Alternate Solutions That Were Considered and Rejected Alternative solutions that were discussed include: Doing nothing - rely on suing the offender: While systems that have accounts have logs that can be mined to locate abusers, it isn't clear that this provides a credible deterrent or defense against the attack described in this document. Systems that don't recognize the situation and take corrective/preventative action are likely to experience failure of a magnitude that precludes retrieval of the records documenting the setup of the attack. (In one scenario, the registrations can occur in a radically different time period than the INVITE transaction. The INVITE request Sparks, et al. Standards Track [Page 17] RFC 5393 Amplification Vulnerability in SIP December 2008 itself may have come from an innocent). It's even possible that the scenario may be set up unintentionally. Furthermore, for some existing deployments, the cost and audit ability of an account is simply an email address. Finding someone to punish may be impossible. Finally, there are individuals who will not respond to any threat of legal action, and the effect of even a single successful instance of this kind of attack would be devastating to a service provider. Putting a smaller cap on Max-Forwards: The effect of the attack is exponential with respect to the initial Max-Forwards value. Turning this value down limits the effect of the attack. This comes at the expense of severely limiting the reach of requests in the network, possibly to the point that existing architectures will begin to fail. Disallowing registration bindings to arbitrary contacts: The way registration binding is currently defined is a key part of the success of the kind of attack documented here. The alternative of limiting registration bindings to allow only binding to the network element performing the registration, perhaps to the extreme of ignoring bits provided in the Contact in favor of transport artifacts observed in the registration request, has been discussed (particularly in the context of the mechanisms being defined in [OUTBOUND]). Mechanisms like this may be considered again in the future, but are currently insufficiently developed to address the present threat. Deprecate forking: This attack does not exist in a system that relies entirely on redirection and initiation of new requests by the original endpoint. Removing such a large architectural component from the system at this time was deemed too extreme a solution. Don't reclaim breadth: An alternative design of the Max-Breadth mechanism that was considered and rejected was to not allow the breadth from completed branches to be reused (see Section 5.3.3.1). Under this alternative, an introduced request would cause, at most, the initial value of Max-Breadth transactions to be generated in the network. While that approach limits any variant of the amplification vulnerability described here to a constant multiplier, it would dramatically change the potential reach of requests, and there is belief that it would break existing deployments. Sparks, et al. Standards Track [Page 18] RFC 5393 Amplification Vulnerability in SIP December 2008 8. Acknowledgments Thanks go to the implementers that subjected their code to this scenario and helped analyze the results at SIPit 17. Eric Rescorla provided guidance and text for the hash recommendation note. 9. References 9.1. Normative References [RFC2119] Bradner, S., "Key words for use in RFCs to Indicate Requirement Levels", BCP 14, RFC 2119, March 1997. [RFC3261] Rosenberg, J., Schulzrinne, H., Camarillo, G., Johnston, A., Peterson, J., Sparks, R., Handley, M., and E. Schooler, "SIP: Session Initiation Protocol", RFC 3261, June 2002. [RFC5234] Crocker, D. and P. Overell, "Augmented BNF for Syntax Specifications: ABNF", STD 68, RFC 5234, January 2008. 9.2. Informative References [DESIGN] Hilt, V., "Design Considerations for Session Initiation Protocol (SIP) Overload Control", Work in Progress, July 2008. [OUTBOUND] Jennings, C. and R. Mahy, "Managing Client Initiated Connections in the Session Initiation Protocol (SIP)", Work in Progress, October 2008. [REQS] Rosenberg, J., "Requirements for Management of Overload in the Session Initiation Protocol", Work in Progress, July 2008. [RFC1321] Rivest, R., "The MD5 Message-Digest Algorithm", RFC 1321, April 1992. [RFC2543] Handley, M., Schulzrinne, H., Schooler, E., and J. Rosenberg, "SIP: Session Initiation Protocol", RFC 2543, March 1999. [RFC4960] Stewart, R., "Stream Control Transmission Protocol", RFC 4960, September 2007. Sparks, et al. Standards Track [Page 19] RFC 5393 Amplification Vulnerability in SIP December 2008 Authors' Addresses Robert Sparks (editor) Tekelec 17210 Campbell Road Suite 250 Dallas, Texas 75254-4203 USA EMail: RjS@nostrum.com Scott Lawrence Nortel Networks, Inc. 600 Technology Park Billerica, MA 01821 USA Phone: +1 978 288 5508 EMail: scott.lawrence@nortel.com Alan Hawrylyshen Ditech Networks Inc. 823 E. Middlefield Rd Mountain View, CA 94043 USA Phone: +1 650 623 1300 EMail: alan.ietf@polyphase.ca Byron Campen Tekelec 17210 Campbell Road Suite 250 Dallas, Texas 75254-4203 USA EMail: bcampen@estacado.net Sparks, et al. Standards Track [Page 20]