At 23:03:10 UTC on November 13, 2026, Bell Canada's AS577 began originating 1.1.1.0/24, the prefix that holds Cloudflare's 1.1.1.1 public DNS resolver. Cloudflare's status page opened an incident 50 minutes later, titled "Issues with 1.1.1.1 public resolver in Canada", and went on to describe "an active hijack of 1.1.1.0/24" without naming the network. The public BGP archives do: three networks sent RIPE RIS and RouteViews collectors a route for 1.1.1.0/24 that ended in AS577, and the last of them held it until 00:36:46 UTC on November 14, 93 minutes after it first appeared.
The prefix has a ROA that authorizes AS13335 and nobody else, so Bell's announcement was RPKI invalid from its first second, and every network that runs route origin validation (ROV) threw it away. That is why Cloudflare's report names one country. Validation did nothing for the networks that skip it, and it could do nothing for Bell's own subscribers: Bell filters invalid routes it receives from other networks, but this route was its own, and no import filter ever looked at it.
Twelve update messages name the origin
We pulled every update for 1.1.1.0/24 from RIPEstat's bgp-updates endpoint, which covers the RIPE RIS collectors, and parsed the 15-minute update files of 55 RouteViews collectors from 23:00 to 01:15 UTC, 495 files in all. Out of that, eleven RouteViews announcements and one RIS announcement carry AS577 as the origin. Here they are with the duplicates removed, followed by the updates that ended it:
23:03:10 route-views.chicago AS49544 1.1.1.0/24 path 49544 577
23:03:10 route-views2 AS22652 1.1.1.0/24 path 22652 577
23:03:10 route-views7 AS16276 1.1.1.0/24 path 16276 577
23:03:11 rrc00 (RIS) AS22652 1.1.1.0/24 path 22652 577
00:29:44 route-views2 AS22652 1.1.1.0/24 path 22652 13335
00:31:09 route-views7 AS16276 1.1.1.0/24 path 16276 13335
00:36:46 route-views.chicago AS49544 1.1.1.0/24 path 49544 13335 AS22652 is Fibrenoire, a Montreal business ISP whose AS is registered to Videotron. AS16276 is OVH, which sent the route over three separate sessions to route-views7. AS49544 is i3D.net, a game hosting network. All three switched in the same second, and none of them ever sent a withdrawal: each replaced Bell's route with a path straight to Cloudflare when Bell's disappeared.
The announcement was a plain /24. We looked for anything more specific inside 1.1.1.0/24, up to a /32, and found nothing, which sets this apart from the June 2024 incident described further down. Every bad path is exactly two ASes long, so no collector peer was seen passing the route on to anyone else. In RIS, one of the 364 peer sessions that carried the prefix that evening pointed at Bell.
Set against Cloudflare's own log, the route was live for 50 minutes before the incident was opened at 23:53:18, and for 63 minutes before the 00:06:42 update that said Cloudflare had "reached out to the network". The 00:52:36 update reads "The network with the offending route leak has fixed the issue on their end", which matches the data: the last trace was gone 16 minutes earlier. The status page gives the incident 68 minutes, against 93 in the routing table.
A collector only learns each peer's best path, so a network that accepted Bell's route but kept preferring its old one is invisible here. Thirteen networks in RIS hear Bell's prefixes directly, among them TELUS (AS852), EBOX (AS1403), NTT, Arelion and Zayo, and only Fibrenoire's session switched. i3D.net shows both behaviors at once: its RouteViews session in Chicago followed Bell while its six RIS sessions stayed on Cloudflare. Three is the number of networks we can prove, and the real count may be higher.
Same prefix length, same path length, and Bell's route still won
The August hijack of a Hetzner block that we wrote up in September worked by announcing a /24 inside a /16, and longest-prefix match did the rest. This time there was no more-specific to do that work: Cloudflare announces 1.1.1.0/24 and Bell announced 1.1.1.0/24. Before 23:03:10, Fibrenoire reached it over the path 22652 13335, a direct session with Cloudflare, and Bell's offer was 22652 577, also two hops.
A router compares local preference before it looks at path length, and with path length tied it falls back on tiebreakers such as origin type, MED and route age. None of that is visible from outside, so we cannot say why these three preferred Bell. The choice should never have come up in the first place, because one of the two candidates failed origin validation.
Bell validates other people's routes, and that could not save its own customers
Bell is one of the Canadian networks that does this properly. APNIC Labs, which measures whether users can still reach a destination that is only announced with an invalid origin, puts AS577 at 96.38 percent filtering, and Cloudflare's own Is BGP safe yet? list marks Bell "signed + filtering".
Operators attach "reject invalid" to the import policy of their external BGP sessions, and a route the network originates itself, from a static route or a redistribution statement, never crosses one of those sessions. RFC 8481 requires routers to evaluate such routes too, but marking a route invalid and rejecting it are different steps, and the second one is a policy somebody has to write. RFC 8893 describes the missing piece, origin validation on export. Had Bell's border routers applied it, 1.1.1.0/24 would not have left AS577.
Inside Bell, a locally originated route beats anything learned from a peer on most routers, so the people most likely to have lost 1.1.1.1 that Friday evening (the route appeared at 18:03 Eastern) were Bell's own subscribers. This is an inference, since no public collector sits inside AS577 and Cloudflare only said "multiple users in Canada". Still, AS577 is the largest network in APNIC's Canadian sample, with 165,123 of 867,772 measurements, or 19 percent.
How the route came to exist is unknown, and we found no statement from Bell. An accident fits the evidence better than an attack: a single /24, no more-specific, no attempt on 1.0.0.0/24, and a route that was gone half an hour after Cloudflare said it had made contact. Cloudflare's wording moved from "hijack" to "route leak" between two updates. In its write-up of the 2024 incident, Cloudflare blamed part of the prefix's bad luck on the "historical use of 1.1.1.1 by networks in labs or as a testing IP address".
Outside Bell, the route went where nobody checks
APNIC's seven-day figure for Canada on November 19 is 43.79 percent of users behind networks that drop invalid routes. The national number hides a split with almost nothing in the middle.
Fibrenoire scores 0.65 percent and its parent Videotron 0.09 percent. OVH scores 3.41 percent among Canadian users. To be fair to OVH, RoVista's latest measurement, from October 9, rates AS16276 as filtering, on a sample of one test, but its three RouteViews sessions carried an invalid route for 88 minutes all the same. Rogers, Shaw and Cogeco do not feed the public collectors, so we cannot tell whether they took Bell's route. At under 1 percent each, nothing in their configuration would have stopped it.
The transit carriers Bell connects to (Arelion, Lumen, Cogent, NTT, Tata, Zayo and GTT are all listed as filtering by Cloudflare) validate, and that explains the word "Canada" in the incident title. An invalid route can only travel through networks that do not check, and in November 2026 those are mostly at the edge.
1.0.0.1 and IPv6 stayed up because nobody announced them
Cloudflare's 00:06 update said that "1.0.0.1 and public resolver IPv6 prefixes are fully operational". RIS agrees. Between 22:30 and 01:30 UTC, 1.0.0.0/24 and 2606:4700:4700::/48 were announced by AS13335 only, including on the Fibrenoire session that was pointing 1.1.1.0/24 at Bell at the time. A device configured with both resolver addresses lost one of them and, after a timeout, used the other.
Three incidents in 29 months, each with an invalid origin somewhere
On June 27, 2024, Eletronet (AS267613) announced 1.1.1.1/32 and Nova Rede de Telecomunicações (AS262504) leaked 1.1.1.0/24 toward AS1031. The /32 was RPKI invalid, some networks accepted it anyway, and Cloudflare counted "over 300 networks in 70 countries" with degraded or no access.
On July 14, 2025, Cloudflare withdrew the resolver prefixes itself through "an internal configuration error" and the service was gone worldwide for 62 minutes. Two minutes into that outage, an origination of 1.1.1.0/24 by Tata Communications India (AS4755) became visible. It had been sitting behind the legitimate route, and Cloudflare was clear that it "was not the cause of the outage".
November 13 makes three, and each time some network originated address space under 1.1.1.0/24 that it had no ROA for. A ROA protects nobody until a network validates against it, and the limits of RPKI we listed in July assumed at least that validation happens. For more than half of Canadian users it does not.
What to change on a resolver, a client and a border router
If you hand out 1.1.1.1 by DHCP or hardcode it, add 1.0.0.1 and the IPv6 addresses (2606:4700:4700::1111 and 2606:4700:4700::1001). The two IPv4 addresses live in different /24s, and only one of them was announced by Bell. Failover is slow, though. glibc waits 5 seconds by default before it tries the next nameserver, so "reachable" during this incident still meant slow. A second resolver from a different operator removes the shared fate entirely.
Use DNS over TLS or DNS over HTTPS with certificate validation where the client supports it. We have no evidence that anything inside Bell answered queries sent to 1.1.1.1 during those 93 minutes. On plain port 53 a client would have had no way to tell. An encrypted transport pinned to one.one.one.one or cloudflare-dns.com fails closed when the address is answered by someone without Cloudflare's certificate.
If you run an autonomous system, turn on ROV for every external session and test the result from a host inside your network at isbgpsafeyet.com. Then look at the other direction. Put origin validation, or at least a prefix list of your own space, on what you export, because the mistake that reached three other networks on November 13 passed through nobody's import policy on its way out of AS577. And watch your own prefixes in RIS and RouteViews: this route was public for 50 minutes before the incident was opened.
The origin AS of a prefix is one of the fields every Ipregistry lookup returns, along with the network's name, its type and the route as it is announced in BGP. You can try the API with 20,000 free lookups to get started.
Keep users in the loop Ship release notes that get read. Try Noticeable