NETWORK INCIDENTS

Incident file

Breakdown
Cloudflare · BGP routing policy

Removing a prefix-list left Cloudflare's Miami export policies too permissive, and IPv6 routes leaked to some of its peers and providers

Happened ·Published

On 22 January 2026, an automated routing policy change made a Cloudflare router in Miami leak some BGP prefixes. The leak lasted 25 minutes, from 20:25 to 20:50 UTC. The change was meant to stop announcing a Bogotá data center's prefixes from Miami. Removing that prefix-list from the export policies left them too permissive. Only IPv6 traffic was affected. Cloudflare customers saw some impact, and multiple other networks were affected too, because their traffic was funnelled through Cloudflare's Miami data center by mistake.

Incident replay · Watch it rebuilt in the lab

A lab model of the failure pattern, not Cloudflare's systems. Watch on YouTube · Run this lab yourself (setup)

What users saw

  • Congestion on Cloudflare's backbone between Miami and Atlanta. Cloudflare says this would have meant elevated loss for some of its customers' traffic, and higher latency than usual for traffic across those links.
  • Networks whose prefixes were leaked would have had their traffic dropped by firewall filters on Cloudflare's routers, which are designed to accept only traffic for Cloudflare's services and customers. At peak, around 12 Gbps of traffic for those prefixes, arriving at the Miami router, was discarded.

What broke

  1. 19:52 UTCA routine clean-upAt 19:52, a change that would trigger the routing policy bug was merged in Cloudflare's network automation code repository. The change Cloudflare pushed at 20:25 was meant to stop announcing from Miami the prefixes of a Bogotá data center. Cloudflare had been forwarding some IPv6 traffic for Bogotá through Miami, and recent infrastructure upgrades had removed the need.
  2. - prefix-listOne line outThe generated diff removed the same line, prefix-list 6-BOG04-SITE-LOCAL, from a term in nine export policies, including 6-PUBLIC-PEER-OUT and 6-TELIA-ACCEPT-EXPORT. Cloudflare says it "looks innocent at a glance".
  3. route-type internalWhat was leftIn the example Cloudflare shows (6-TELIA-ACCEPT-EXPORT), the term was left matching only "route-type internal", followed by accept. In Junos and Junos EVO that matches any route that isn't external, including routes learned over internal BGP (IBGP), which is what happened here.
  4. 20:25 UTCThe leakAutomation ran the change on one Miami edge router. Every IPv6 prefix Cloudflare redistributes internally across its backbone was accepted and advertised to all its BGP neighbours in Miami, so routes learned from some of its peers were sent on to some of its peers and providers. Cloudflare calls this very similar to its 2020 outage.
  5. RESULTA peer's route sent upstreamCloudflare (AS13335) took a prefix it had received from its peer Meta (AS32934) and advertised it to Lumen (AS3356), one of its upstream transit providers. Routes from a peer are only meant to be passed on to customers. The updates can still be seen in route-collector data (MRT files), with tools such as monocle or RIPE BGPlay.

What makes it a route leak

A route leak is a network telling the broader Internet to send it traffic it isn't supposed to forward: technically, an autonomous system (AS) appearing unexpectedly in an AS path. Once a route has been sent to a peer or provider, it should only be passed on to customers, never to another peer or provider.

In Cloudflare's general description, the leaking network is often not prepared to handle the amount of traffic it receives, and may not even have firewall filters configured to accept all of it. By the definitions in RFC 7908, Cloudflare caused a mixture of Type 3 and Type 4 route leaks.

Timeline (UTC)

  1. The change that will trigger the routing policy bug is merged in the network automation code repository.
  2. Automation runs on one Miami edge router; unexpected advertisements go to BGP transit providers and peers. Impact starts.
  3. The network team begins investigating the unintended route advertisements from Miami.
  4. An incident is raised to coordinate the response.
  5. A network operator reverts the bad configuration change by hand, and automation is paused for the router so it can't run again. Impact stops.
  6. The change that triggered the leak is reverted in the code repository.
  7. Operators confirm the automation is healthy to run again on the Miami router, without the routing policy bug.
  8. Automation is unpaused on the Miami router.

The networking lesson

  • Catch empty policy terms before they ship. Cloudflare is patching the failure in its routing policy automation, and adding automatic routing policy evaluation to its CI/CD pipelines that looks specifically for empty or erroneous policy terms. It also plans to improve early detection of problems with network configurations and the effects of automated changes.
  • Reject peer and provider routes on the way out. It is adding BGP community-based safeguards that explicitly reject, on external export policies, routes received from providers and peers.
  • Guardrails that don't depend on policy. Cloudflare is validating vendors' implementations of RFC 9234 (BGP roles and the Only-to-Customer attribute) before rolling it out, calling it the only way, independent of routing policy, to prevent route leaks caused at the local AS. It also encourages long-term adoption of RPKI ASPA, which would let networks automatically reject routes with anomalous AS paths.
  • Deleting a match condition widens a term. A term that ends in accept accepts everything its remaining conditions match. When a change removes a condition, review what the whole term now matches, not just the line that went. our analysis
  • Reverted and paused together. At 20:50 an operator reverted the change by hand, and automation was paused for the router so it couldn't run again.
  • A manual revert needs the automation stopped too. If automation generated the bad config, its next run can put it straight back. Stop the automation in the same step as the revert. our analysis

Source: Route leak incident on January 22, 2026, Cloudflare (official post-mortem). Facts, times and causes are as the source states them; anything marked our analysis is ours.

More incident breakdowns

  • AWS, 20 Oct 2025Two copies of AWS's DNS automation raced, and DynamoDB's regional endpoint was left with an empty DNS record
  • Cloudflare, 18 Nov 2025A database permissions change doubled a config file, and Cloudflare's proxy fell over