The Use of Maxlength in the RPKI

The RPKI uses Route Origin Authorizations (ROAs) to create a trusted mapping from an IP prefix to a set of autonomous systems (ASes) that are authorized to originate this prefix. Each ROA contains a set of IP prefixes, and an AS number of an AS authorized originate all the IP prefixes in the set . Each ROA is cryptographically signed by the party that is authorized to allocate the set of IP prefixes. The RPKI also supports a maxLength attribute. According to , "When present, the maxLength specifies the maximum length of the IP address prefix that the AS is authorized to advertise." Thus, rather than requiring the ROA to explictly list each prefix the AS is authorized to originate, the maxLength attribute provides a shorthand that authorizes an AS to originate a set of IP prefixes. However, measurements of current RPKI deployments have found that use of the maxLength in ROAs tends to lead to security problems. Specifically, as of June 2017, 84% of the prefixes specified in ROAs that use the maxLength attribute, are vulnerable to a forged-origin subprefix hijack. The forged-origin subprefix hijack can be launched against any IP prefix that is authorized in ROA but is not originated in BGP. The impact of such an attack is the same as standard subprefix hijack on an IP prefix that is unprotected by a ROA in the RPKI. For this reason, this document recommends that, whenever possible, operators SHOULD use "minimal ROAs" that include only those IP prefixes that are actually originated in BGP, and no other prefixes. Operators SHOULD also avoid using the maxLength attribute in their ROAS whenever possible. One ideal place to implement these recommendations is in the user interfaces for configuring ROAs. The recommendations in this document clarify and extend the following recommendation from : One advantage of minimal ROA length is that the forged origin attack does not work for sub-prefixes that are not covered by overly long max length. For example, if, instead of 10.0.0.0/16-24, one issues 10.0.0.0/16 and 10.0.42.0/24, a forged origin attack cannot succeed against 10.0.666.0/24. They must attack the whole /16, which is more likely to be noticed because of its size. This best current practice requires no changes to the RPKI specification and will not increase the number of signed ROAs in the RPKI, because ROAs already support lists of IP prefixes .

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 .

It is assumed that the reader understands BGP , the RPKI Route Origin Authorizations (ROAs) , RPKI-based Prefix Validation , and BGPSEC .

The forged-origin subprefix hijack is relevant to a scenario in which (1) the RPKI is deployed, and (2) routers use RPKI origin validation to drop invalid routes , but (3) BGPSEC is not deployed. We describe the forged-origin subprefix hijack using a running example. Consider the IP prefix 168.122.0.0/16 which is allocated to an organization that also operates AS 111. In BGP, AS 111 originates the IP prefix 168.122.0.0/16 as well as its subprefix 168.122.225.0/24. Therefore, the RPKI should contain a ROA authorizing AS 111 to originate these two IP prefixes. That is, the ROA should be ROA:(168.122.0.0/16,168.122.225.0/24, AS 111) This ROA is "minimal" because it includes only those IP prefixes that AS 111 originates in BGP, but no other IP prefixes. Now suppose an attacking AS 666 originates a BGP announcement for a subprefix 168.122.0.0/24. This is a standard "subprefix hijack". In the absence of the minimal ROA above, AS 666 could intercept traffic for the addresses in 168.122.0.0/24. This is because routers perform a longest-prefix match when deciding where to forward IP packets, and 168.122.0.0/24 originated by AS 666 is a longer prefix than 168.122.0.0/16 originated by AS 111. However, the minimal ROA renders AS 666's BGP announcement invalid, because (1) this ROA "covers" the attacker's announcement (since 168.122.0.0/24 is a subprefix of 168.122.0.0/16), and (2) there is no ROA "matching" the attacker's announcement (there is no ROA for AS 666 and IP prefix 168.122.0.0/24) . If routers ignore invalid BGP announcements, the minimal ROA above ensures that the subprefix hijack will fail. Now suppose that instead the "minimal ROA" was replaced with a "loose ROA" that used maxLength as a shorthand for set of IP prefixes that AS 111 is authorized to originate. The "loose ROA" would be: ROA:(168.122.0.0/16-24, AS 111) This "loose ROA" authorizes AS 111 to originate any subprefix of 168.122.0.0/16, up to length /24. That is, AS 111 could originate 168.122.225.0/24 as well as all of 168.122.0.0/17, 168.122.128.0/17, ..., 168.122.255.0/24 but not 168.122.0.0/25. However, AS 111 only originates two prefixes in BGP: 168.122.0.0/16 and 168.122.255.0/24. This means that all other prefixes authorized by the "loose ROA" (for instance, 168.122.0.0/24), are vulnerable to the following forged-origin subprefix hijack [,]: The hijacker AS 666 sends a BGP announcement "168.122.0.0/24: AS 666, AS 111", falsely claiming that AS 666 is a neighbor of AS 111 and falsely claiming that AS 111 originates the IP prefix 168.122.0.0/24. In fact, the IP prefix 168.122.0.0/24 is not originated by AS 111. The hijacker's BGP announcement is valid according the RPKI, since the ROA (168.122.0.0/16-24, AS 111) authorizes AS 111 to originate BGP routes for 168.122.0.0/24. Becaue AS 111 does not actually originate a route for 168.122.0.0/24, the hijacker's route is the *only* route to the 168.122.0.0/24. Longest-prefix-match routing ensures that the hijacker's route to the subprefix 168.122.0.0/24 is always preferred over the legitimate route to 168.122.0.0/16 originated by AS 111. Thus, if the hijacker's route propagates through the Internet, the hijacker will intercept traffic destined for IP addresses in 168.122.0.0/24. The forged origin *subprefix* hijack would have failed if the "minimal ROA" described above was used instead of the "loose ROA". If the "minimal ROA" had been used instead, the attacker would be forced to launch a forged origin *prefix* hijack in order to attract traffic, as follows: The hijacker AS 666 sends a BGP announcement "168.122.0.0/16: AS 666, AS 111", falsely claiming that AS 666 is a neighbor of AS 111. This forged-origin *prefix* hijack is significantly less damaging than the forged-origin *subprefix* hijack. With a forged-origin *prefix* hijack, AS 111 legitimately originates 168.122.0.0/16 in BGP, so the hijacker AS 666 is not presenting the *only* route to 168.122.0.0/16. Moreover, the path originated by AS 666 is one hop longer than the path originated by the legitimate origin AS 111. As discussed in , this means that the hijacker will attract less traffic than he would have in the forged origin *subprefix* hijack, where the hijacker presents the *only* route to the hijacked subprefix. In sum, a forged-origin subprefix hijack has the same impact as a regular subprefix hijack. A forged-origin *subprefix* hijack is also more damaging than than forged-origin *prefix* hijack. Any ROA that is not minimal is vulnerable to forged-origin subprefix hijack.

Network measurements from June 1, 2017 show that 12% of the IP prefixes authorized in ROAs have a maxLength longer than their prefix length. The vast majority of these (84%) of these are vulnerable to forged-origin subprefix hijacks. Even large providers are vulnerable to these attacks. See for details. These measurements suggest that operators commonly misconfigure the maxLength attribute, and unwittingly open themselves up to forged-origin subprefix hijacks.

Operators SHOULD avoid using the maxLength attribute in their ROAs. Operators SHOULD use "minimal ROAs" whenever possible. A minimal ROA contains only those IP prefixes that are actually originated by an AS in BGP, and no other IP prefixes. (See for an example.) This practice requires no changes to the RPKI specification and will not increase the number of signed ROAs in the RPKI, because ROAs already support lists of IP prefixes . See also for further discussion of why this practice will have minimal impact on the performance of the RPKI ecosystem.

Sometimes, it is not possible to use a "minimal ROA", because an operator wants to issue a ROA that includes an IP prefix that is sometimes (but not always) originated in BGP. In this case, the ROA SHOULD include (1) the set of IP prefixes that are always originated in BGP, and (2) the set IP prefixes that are sometimes, but not always, originated in BGP. The ROA SHOULD NOT include any IP prefixes that the operator knows will not be originated in BGP. Whenever possible, the ROA SHOULD also avoid the use of the maxlength attribute. We now extend our running example to illustrate one situation where where it is not possible to issue a minimal ROA. Suppose AS 111 has a contract with a DDoS mitigation service provider that holds AS 222. The DDoS mitigation service is contracted to protect all IP addresses covered by 168.122.0.0/22. When a DDoS attack is detected, AS 222 immediately originates 168.122.0.0/22, thus attracting all the DDoS traffic to itself. The traffic is scrubbed at AS 222 and then and sent back to AS 111 over a backhaul data link. Notice that, during a DDoS attack, the DDoS mitigation service provider AS 222 originates a /22 prefix that are longer than than AS 111's /16 prefix, and so all the traffic that normally goes to AS 111 goes to AS 222 instead. First, suppose the RPKI only had the minimal ROA for AS 111, as described in . But, if there is no ROA authorizing AS 222 to announce the /23 prefix, then the traffic-scrubbing scheme would not work. That is, if AS 222 originates the /22 prefix in BGP during a DDoS attack, the announcement would be invalid . Instead, the RPKI should have two ROAs: one for AS 111 and one for AS 222. ROA:(168.122.0.0/16,168.122.225.0/24, AS 111) ROA:(168.122.0.0/22, AS 222) Neither ROA uses the maxLength attribute. But, the second ROA is not "minimal" because it contains a /22 prefix that is not originated by anyone in BGP during normal operations. The /22 prefix is only originated by AS 222 as part of its DDoS mitigation service during a DDoS attack. Notice, however, that this scheme does not come without risks. Namely, all of the IP addresses in 168.122.0.0/22 are vulnerable to a forged-origin subprefix hijack during normal operations, when the /22 prefix is not originated. (The hijacker AS 666 would send the BGP announcement `168.122.0.0/22: AS 666, AS 222'', falsely claiming that AS 666 is a neighbor of AS 222 and falsely claiming that AS 222 originates 168.122.0.0/22.) In some situations, the DDoS mitigation service at AS 222 might want to limit the amount of DDoS traffic that it attracts and scrubs. Suppose that a DDoS attack only targets IP addresses in 168.122.0.0/24. Then, the DDoS mitigation service at AS 222 only wants to attract the traffic destinated for the /24 prefix that is under attack, but not the entire /22 prefix. To allow for this, the RPKI should have two ROAs: one for AS 111 and one for AS 222. ROA:(168.122.0.0/16,168.122.225.0/24, AS 111) ROA:(168.122.0.0/22-24, AS 222) The second ROA uses the maxLength attribute because it is designed to explicitly enable AS 222 to originate *any* /24 subprefix of 168.122.0.0/22. As before, the second ROA is also not "minimal" because it contains prefixes that are not originated by anyone in BGP during normal operations. As before, all of the IP addresses in 168.122.0.0/22 are vulnerable to a forged-origin subprefix hijack during normal operations, when the /22 prefix is not originated. The use of maxLength in this second ROA also comes with an additional risk. Consider a DDoS attack that causes the DDoS mitigation service at AS 222 to originates prefix 168.122.0.0/24. It follows that all *other* /24 prefixes covered by the /22 prefix (i.e., 168.122.1.0/24, 168.122.2.0/24, 168.122.3.0/24) are all vulnerable to a forged-origin subprefix attacks during the DDoS attack.

Note to RFC Editor: if this document does not obsolete an existing RFC, please remove this appendix before publication as an RFC. 00 - New document. 01 - Updated network measurements. Updated citation to . Editorial changes to clarify difference between "minimal ROA" recommendation and "avoid maxLength" recommendation. Updated example in .

This document would not be possible without the work of Omar Sagga (Boston University).