Short · Network security · 18 min read

Does RPKI Stop BGP Route Leaks?

RPKI route-origin validation does not, by itself, stop BGP route leaks. It checks whether the origin AS claimed by a route is authorized for the announced prefix and length. A route can pass that check while a network propagates it to a neighbor that should never receive it under their routing relationship. You need origin checks and appropriate import/export controls, with evidence that both are operating.

The distinction here concerns ROA-based origin validation. RPKI is supporting infrastructure, not a synonym for one validation procedure. This paper does not claim that every mechanism using that infrastructure has the same limits; it examines the common assurance gap between an authorized origin and permitted onward propagation.

I would challenge any assurance statement that reduces those different properties to “we have RPKI.” Creating Route Origin Authorizations, validating received routes, rejecting invalid origins and preventing inappropriate transit are distinct activities. An organization can perform one well while leaving another unimplemented. The practical question is which unwanted announcement each control will actually reject.

This paper develops a proposed routing-security assessment using twelve original standards and government publications checked on October 29, 2026. Its acceptance record, hypothetical cases and testing sequence are original synthesis, not an empirically validated universal design. No router was accessed and no route was advertised. Named standards describe mechanisms; actual support, policy and failure behavior require verification in the operator's environment.

Begin with the claim that origin validation actually makes

An autonomous system, or AS, is a routing domain identified by an AS number. A Route Origin Authorization, or ROA, expresses authorization concerning an address prefix and an originating AS. A relying-party validator checks RPKI material and provides validated information that routers can use. That is different from asking whether every network carrying an announcement was supposed to offer transit for it.

RFC 6811 defines three origin-validation states against the available validated ROA payloads. Valid means at least one payload matches the prefix, permitted maximum length and origin AS. Invalid means at least one payload covers the prefix but none matches. NotFound means no payload covers it (Mohapatra et al., 2013, §2). A conflicting covering entry does not override an existing match merely because it was examined first.

These names are technical states, not general security verdicts. Valid does not mean that the traffic will reach the intended application or that the entire path complies with business relationships. Invalid does not establish malicious intent; an authorized operational change can be inconsistent with published authorization. NotFound describes the absence of covering validation data in the view being used, not proof that the announcement is safe or hostile.

I would preserve the route, the relevant validated records and the observation time when explaining a result. A dashboard color without those inputs cannot distinguish an unexpected origin, an excessive prefix length or an incomplete validation view. That distinction matters because the corrective action differs: repair an authorization, correct an announcement, investigate a policy violation, or diagnose the validation-data path.

A legitimate origin can sit at the end of a leaked route

RFC 7908 defines a route leak as propagation beyond the intended scope, including violations of the policies associated with the networks along the path. Its relationship model includes customers, transit providers and peers (Sriram et al., 2016, §2). The origin can remain unchanged while another AS makes an inappropriate onward advertisement. Origin validation has no relationship claim to evaluate in that situation.

Consider a simplified example. Customer C buys transit from providers P1 and P2. C learns a route to Origin O through P1. If C then advertises that learned route to P2 as though C were offering transit, P2 may receive a route whose origin is still O. Where O is correctly authorized for that prefix, the origin check can remain Valid. The defect lies in the propagation relationship, not the origin identifier.

This is a conceptual counterexample, not a report of a measured incident. It demonstrates why one successful validation cannot establish a different property. Removing the invalid-origin routes from a routing table does not remove every route that should be rejected for another reason. A leak can affect a single legitimate prefix; it does not need to produce an enormous announcement burst to matter.

Start the assessment by stating which role your network occupies on each session and what routes that relationship permits you to send and receive. A multihomed enterprise that does not sell transit needs a different export boundary from a provider serving downstream customers. The contract and actual topology need to agree before an engineer can turn them into useful route policy.

Validation state and filtering are separate controls

A router can calculate a validation state without rejecting anything. RFC 8481 clarifies that state evaluation and policy action are separate: absent specific operator configuration, origin-validation policy must not be applied automatically (Bush, 2018, §§4–5). “The router supports ROV” therefore does not establish that an invalid route is ineligible under the installed policy.

I would inspect the effective import policy, its ordering and its attachment to the relevant sessions and address families. Then compare the classification result with route eligibility, selected routing information and forwarding behavior. A monitoring screen that labels a route Invalid can be perfectly accurate while the organization still forwards through it. Those observations answer different questions and should appear separately in acceptance evidence.

Merely reducing preference is also not equivalent to rejection. RFC 7115 warns that a more-specific invalid announcement can still attract forwarding through longest-prefix matching even when its routing preference is reduced (Bush, 2014, §5). The distinction is between choosing among paths to a prefix and choosing the most-specific matching destination prefix. A policy review that confuses those decisions can produce an ineffective mitigation.

Use an explicit policy for Valid, Invalid and NotFound, including any bounded exceptions. Don't silently translate NotFound into Invalid. If a temporary exception permits an otherwise rejected route, bind it to the precise prefix, origin, purpose and review condition. Record the capability it restores and the protection it gives up. An exception that disables validation globally to repair one reachability problem changes far more than the original incident requires.

Publish authorization that matches the routes you intend to originate

ROV depends on the authorization set being accurate. Inventory the prefixes you originate, their intended origin ASes, planned more-specific announcements and legitimate alternative origins. Include failover and mitigation arrangements. The production plan and the recovery plan need compatible authorization; a backup origin that becomes Invalid during an outage is a recovery dependency that should have been visible before the outage.

The maxLength field deserves deliberate review. RFC 9319 recommends minimal ROAs that authorize the prefixes actually originated, and limiting unnecessary use of maxLength to reduce the forged-origin attack surface (Gilad et al., 2022). A broad maximum is not a harmless convenience. It authorizes a larger set of prefix lengths, even if you currently advertise only the aggregate.

As a hypothetical illustration, an authorization for a /24 with a maximum length of /24 does not authorize its /25 subprefix merely because the origin AS is unchanged. A maximum of /25 changes that result. This example does not recommend either setting: the right authorization depends on which announcements are genuinely intended. Document the reason for permitted deaggregation instead of expanding the range until every current test turns green.

Changes need sequencing. Before moving origination or adding a planned more-specific route, determine when the required authorization will be visible to the validation views that matter. After the transition, remove obsolete permissions when the legitimate requirement ends. Avoid assuming a control-panel update instantly changes every router's data. Verify observed state at the boundaries in scope and retain a rollback plan consistent with both the old and new announcement arrangements.

Build neighbor policy from the relationship, not the RPKI label

For a customer-facing session, ask which prefixes the customer and its authorized downstream network may originate. For a transit-facing session, ask which of your routes the provider should receive. For a peer, establish the intended exchange scope. These questions describe permitted propagation. ROV supplies an additional origin check; it does not supply the missing relationship policy.

NIST SP 800-189's customer-filtering guidance distinguishes cases with reliable knowledge of the customer's prefixes and customer cone from cases without it. Where that knowledge is reliable, it recommends accepting the known customer/cone prefixes and rejecting others (Sriram & Montgomery, 2019, §4.5.3). The qualification matters: an incomplete inventory is not a defensible basis for pretending the allowlist is complete.

I would version the intended prefix scope alongside the configuration derived from it. Identify how customer changes, additional downstream customers, migrations and emergency service arrangements update that scope. Compare generated policy with the approved source rather than treating successful configuration compilation as evidence that the business relationship was represented correctly. Automation can faithfully install the wrong assumption at scale.

An explicit default boundary helps prevent missing policy from becoming unrestricted exchange. RFC 8212 specifies that eBGP routes without an explicit import policy are not eligible for the decision process and routes without an explicit export policy are not added to the relevant outbound routing information. It also permits implementations to offer a configuration option that deviates from that behavior (Mauch et al., 2017, §3). Inspect actual behavior rather than inferring it from a product's standards-support list.

Use BGP Roles and OTC for the relationship they encode

RFC 9234 adds BGP Roles and the Only to Customer, or OTC, attribute to support route-leak prevention and detection. Roles allow the neighbors to establish their relationship; OTC constrains subsequent propagation. Among its rules, a route received with OTC from a customer is ineligible, and a route already carrying OTC must not be propagated to providers, peers or route servers (Azimov et al., 2022, §§4–5).

That mechanism addresses a question origin validation leaves open: how a learned route may move through the relationship structure. It should be assessed as a complementary control. A correctly classified origin and a correctly enforced propagation constraint can both be necessary. Passing either check alone does not excuse failure of the other.

Scope still matters. The RFC's procedures target IPv4 and IPv6 unicast, and it excludes configuring Roles on a session with a complex peering relationship that does not fit the ordinary role model (Azimov et al., 2022, §§5–6). Do not force a complex commercial arrangement into a convenient role label merely to claim deployment. Separate the relationships where practical or document the applicable per-prefix policy and its limitations.

For acceptance, inspect both sides' intended roles, implementation support and resulting route handling. Test a permitted announcement as well as an announcement that should be ineligible. A session that fails to establish is not evidence that the desired operational design works. Likewise, observing an OTC attribute does not establish that every relevant import and export path applies the intended rules.

Do not replace one broad assurance claim with another

BGPsec addresses path authenticity through signatures along a BGP update's AS path. RFC 8205 describes the resulting assurance about the update's propagation and explicitly says it does not guarantee that data packets follow the indicated path (Lepinski & Sriram, 2017, §8.1). That is a useful, stronger claim about a different object; it is not a universal statement that the service or its traffic is secure.

I would keep three questions visible: is the claimed origin authorized, is the update path authentic under the mechanism being used, and is propagation consistent with the intended relationship? A network may intentionally sign an advertisement produced by a mistaken policy. Cryptographic authenticity cannot substitute for deciding whether that policy was appropriate.

These distinctions also prevent overpromising to business leaders. Route controls can reduce particular routing risks without proving confidentiality, application authenticity or availability. An encrypted application connection and a routing-policy boundary address different failure mechanisms. Neither should be described as replacing the other simply because both appear in a diagram labeled “network security.”

Treat validation data as an operational dependency

The router's answer depends on its available validation data. Separate repository and relying-party health, cache-to-router delivery, local route classification and policy action. A connected session is only one observation in that chain. It does not establish that the received data are the intended, sufficiently current view or that the router reevaluated affected routes correctly.

The version 1 RPKI-to-Router protocol defines refresh, retry and expiry behavior. Its Expire Interval limits how long a router may retain the cache's data when it cannot complete a successful subsequent query (Bush & Austein, 2017, §6). This is a statement about the named protocol, not a claim that every product or later protocol version uses identical defaults.

A loss of covering data can change validation results. If an announcement previously classified Invalid loses its only covering payload, its result may become NotFound; other surviving covering payloads can change that outcome. Under a policy that rejects Invalid but accepts NotFound, the data-loss event can alter eligibility without a new route announcement. That conditional consequence belongs in the failure review, not in an unsupported claim that every validator outage automatically opens every route.

Record what the implementation does when one cache fails, all configured caches become unavailable, data expire, or data disagree. Multiple validators can improve resilience, but shared repositories, software or management paths can create common dependencies. Inspect the actual combination behavior. Two configured endpoints are not proof that the intended data view survives either endpoint's failure.

Test changes in validation state as well as new announcements

A static demonstration is incomplete. A route may remain advertised while the authorization data change. The implementation must update the relevant classification and policy result, including reconsidering routes previously excluded. Otherwise a corrected authorization can remain operationally ineffective, or a newly invalid route can remain selected longer than the acceptance requirement permits.

RFC 9324 updates RFC 8481 to address reevaluation without imposing unnecessary route-refresh work on neighbors. It requires retaining a full inbound routing view or at least the routes dropped because of ROV for potential future evaluation, with the described implementation controls (Bush et al., 2022, §§4–5). The operational issue is preserving enough candidate information to reevaluate when validation data change.

I would test both directions: an accepted route becomes invalid under revised data, and a previously rejected route becomes valid after a legitimate correction. Observe classification, eligibility and the selected route separately. Record which data revision triggered the transition and whether the change caused unexpected refresh activity or resource pressure. A label changing on one screen does not establish that forwarding changed as intended.

Set the acceptable transition time from the service's needs and measured implementation behavior. This paper supplies no universal convergence deadline. Include the validator's update cycle, transport, router processing and any operational hold points. A number measured in one small laboratory cannot be presented as a guarantee for every route count, failure mode or hardware platform.

Use volume limits without mistaking them for authorization

A maximum-prefix control can help contain an unexpected routing-table expansion. RFC 7454 recommends limits appropriate to the peer or upstream, with growth headroom and consideration of the route volume expected from that relationship (Durand et al., 2015, §8). That is a quantity control. It does not decide whether each route within the permitted quantity belongs there.

A one-prefix leak can stay below every sensible maximum. Conversely, legitimate growth can cross a poorly maintained limit and interrupt a session. I would therefore document what the limit protects, how warning and enforcement differ, how headroom is reviewed, and what action follows a threshold breach. A permanently inflated limit avoids one operational nuisance while abandoning the reason for having the control.

Combine volume observations with relationship and origin evidence. If a customer suddenly advertises many additional prefixes, investigate both the quantity and the authorization of the new set. If only one critical destination changes, inspect it even when total route count barely moves. Aggregate stability cannot establish that the particular route supporting an essential service is acceptable.

Build an acceptance matrix that contains a valid-origin leak

The acceptance exercise should challenge the distinction that motivated the assessment. Use an isolated lab or approved provider test arrangement with documentation prefixes and controlled AS identities; never inject a test leak into public routing. The following matrix describes expected outcomes under an illustrative policy that rejects Invalid origins and prohibited customer transit. It reports no executed tests.

Hypothetical input or changeExpected bounded observation
Intended origin, permitted length and permitted relationshipRoute remains eligible under the complete policy
Covered prefix with an unauthorized claimed originOrigin state is Invalid; configured rejection is observed
Correct origin but length beyond every matching authorizationOrigin state is Invalid; length cause is distinguishable
No covering payloadNotFound is recorded and the explicit NotFound policy applies
Valid-origin route learned from a provider and offered as prohibited customer transitRelationship control rejects the leak despite Valid origin state
A required import or export policy is missingIntended default behavior is observed without unintended propagation
Applicable OTC-bearing route received from a customerExpected ineligibility is observed under the implemented RFC 9234 behavior
Relevant validation data change while BGP advertisement remains unchangedClassification and eligibility are reevaluated in both directions
Cache loss followed by expiryResulting data view, classification and policy consequence match the documented failure design

Keep the positive case. A control that blocks every advertisement can pass a negative-only demonstration while making the service unusable. For each negative case, keep unrelated conditions valid so that the intended mechanism is responsible for rejection. A malformed update does not test whether a well-formed, valid-origin leak would be refused.

Retain the sanitized topology, intended relationships, route inputs, authorization snapshot, configuration version, observation points and expected versus actual outcomes. A test proves behavior for that scope and state. It does not establish global deployment, universal leak prevention or absence of implementation defects. Repeat relevant cases when relationships, routing software, validators, authorization or policy generation change.

Work through a multihomed enterprise decision

Consider fictional Harrowmere, a company with two transit providers and no business requirement to carry traffic between them. Its own prefix has an accurate ROA, and both routers reject Invalid origins. Management has been told that routing security is complete because the RPKI dashboard is green.

In a design review, the team finds that one transit-facing export policy permits learned routes beyond Harrowmere's intended advertisements. A route learned from Provider A could be sent to Provider B with its original, authorized origin intact. The origin check would not address that propagation error. The required correction is to restore the export boundary and establish appropriate complementary controls with the providers, not to change the unrelated origin's ROA.

A second finding concerns recovery. Harrowmere's planned mitigation service would originate a more-specific prefix under a different authorized arrangement, but the published authorization does not cover the planned state. The response is to reconcile the legitimate mitigation design, authorization and provider acceptance before relying on it. Broadening every ROA indefinitely would hide the immediate mismatch by creating unnecessary authorization elsewhere.

The team would then test ordinary operation, the prohibited transit case and the planned recovery announcement in its authorized environment. It would separately test validation-data failure. These hypothetical actions demonstrate distinct decisions: prevent an unwanted export, authorize a wanted recovery route, and define behavior when evidence is unavailable. No real organization, route experiment or measured outcome is implied.

Ask providers for evidence that matches the service you buy

An enterprise using provider-managed routing may not control the relevant import policy or validators. That does not make a generic assurance statement sufficient. Ask which boundaries the provider operates, which traffic and sessions are covered, and how the provider distinguishes origin validation from customer and peer filtering. A service description should identify the scope it actually supports.

Useful evidence can include a sanitized policy description, the treatment of validation states, relevant role/filter support, change procedures and a bounded demonstration or independently reviewed result. Request the exception and outage behavior as well as steady-state behavior. A provider may reasonably limit disclosure of sensitive topology or configuration; that limitation should lead to appropriately scoped evidence, not an invented claim of independent verification.

Keep responsibility for your own authorizations, intended announcements and recovery changes explicit. A provider cannot reliably infer every future failover plan from current advertisements. Coordinate changes affecting permitted prefixes and origins through the agreed operational process. Evidence that a provider filters one customer session does not establish that every other Internet network applies the same filtering.

Balance reachability, uncertainty and operational risk

The strongest objection to strict rejection is that incorrect authorization can disrupt legitimate connectivity. That is a real operational concern. Use staged observation, reconcile legitimate routes and define narrow exceptions before enforcement. But do not describe a monitoring-only deployment as protection from invalid announcements. State what it observes and which policy actions remain pending.

Another objection is that incomplete adoption limits the value of a local control. It does limit the assurance claim, especially for paths outside your administrative boundary. It does not make local import and export policy irrelevant. Distinguish preventing your network from accepting or creating a particular error from proving that nobody else can create one. The latter is a much broader claim than a local configuration can support.

Routing is also only one recovery dependency. The DNS recovery analysis examines how resolver, authority and application behavior affect service restoration. The cloud egress analysis examines whether a permitted transfer reaches an authorized recipient. Neither question is resolved by a Valid origin label, and this article does not duplicate those application and information-flow assessments.

I would report separate evidence for authorization accuracy, validation enforcement, relationship policy and failure handling. Combining those into a single percentage would conceal different denominators and untested conditions. “Nine cases behaved as expected in this lab” is a bounded observation if actually measured; it is not a percentage of Internet routing secured. Here, even those cases remain proposed rather than executed.

What a defensible routing-security claim sounds like

The standards support distinct mechanisms, not a guarantee that any particular network has deployed them correctly. NIST's cited publication is the 2019 final version; a later revision draft is not treated here as an adopted requirement. The named RFCs define the referenced behavior, including updates relevant to origin-state evaluation. Platform support, newer changes and effective operator policy still require review before deployment.

The defensible statement is specific: these prefixes have the intended origin authorizations; these sessions apply the stated validation and propagation policies; these controlled cases produced the recorded results; these dependencies and exceptions remain. That statement is more useful than “RPKI stops route leaks” because it tells the next reviewer what can fail and where to look.

If you need to establish that evidence, request a routing-security and resilience assessment. Start with one critical connectivity service, its intended announcements and its provider relationships. I can help separate origin authorization from route-leak controls, review the recovery assumptions and define a bounded acceptance exercise without treating a green validation badge as proof that every path is trustworthy.

References

  • Azimov, A., Bogomazov, E., Bush, R., Patel, K., & Sriram, K. (2022). Route leak prevention and detection using roles in UPDATE and OPEN messages (RFC 9234). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc9234
  • Bush, R. (2014). Origin validation operation based on the Resource Public Key Infrastructure (RPKI) (RFC 7115). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc7115
  • Bush, R. (2018). Clarifications to BGP origin validation based on Resource Public Key Infrastructure (RPKI) (RFC 8481). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc8481
  • Bush, R., & Austein, R. (2017). The Resource Public Key Infrastructure (RPKI) to Router Protocol, version 1 (RFC 8210). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc8210
  • Bush, R., Patel, K., Smith, P., & Tinka, M. (2022). Policy based on the Resource Public Key Infrastructure (RPKI) without route refresh (RFC 9324). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc9324
  • Durand, J., Pepelnjak, I., & Doering, G. (2015). BGP operations and security (RFC 7454). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc7454
  • Gilad, Y., Goldberg, S., Sriram, K., Snijders, J., & Maddison, B. (2022). The use of maxLength in the Resource Public Key Infrastructure (RPKI) (RFC 9319). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc9319
  • Lepinski, M., & Sriram, K. (Eds.). (2017). BGPsec protocol specification (RFC 8205). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc8205
  • Mauch, J., Snijders, J., & Hankins, G. (2017). Default external BGP (EBGP) route propagation behavior without policies (RFC 8212). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc8212
  • Mohapatra, P., Scudder, J., Ward, D., Bush, R., & Austein, R. (2013). BGP prefix origin validation (RFC 6811). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc6811
  • Sriram, K., & Montgomery, D. (2019). Resilient interdomain traffic exchange: BGP security and DDoS mitigation (NIST SP 800-189). National Institute of Standards and Technology. https://doi.org/10.6028/NIST.SP.800-189
  • Sriram, K., Montgomery, D., McPherson, D., Osterweil, E., & Dickson, B. (2016). Problem definition and classification of BGP route leaks (RFC 7908). Internet Engineering Task Force. https://www.rfc-editor.org/rfc/rfc7908