
The DNSSEC root changed its key on 11 October: what broke, hour by hour, and which environments already trust the new one
On 11 October 2026 the DNS root started signing with KSK-2024. I watched ten resolvers every five minutes. For example.com, counting from the first round with the new key, Ubuntu 24.04's systemd-resolved started failing after 20 minutes, the unbounds holding only the old key between seven and a half and thirteen and a half hours later, Pi-hole v5 stayed up and Yandex stopped marking its answers as validated. Before the switch I measured which distros, Docker images, routers and resolvers already trust it.
On Sunday 11 October the DNS root changed the key it signs with; my vigil first saw it signing with the new one at 14:04 UTC. It is the second time this happens: the first was in October 2018. Before the switch I expected almost nobody to notice: according to APNIC, most users sit behind resolvers that do not validate DNSSEC, and Cloudflare, Quad9 and AdGuard, which do validate, already trust the new key. What interested me was the rest: what I run myself and what people run at home, a Pi-hole, an OpenWrt router, an unbound in a container.
So I measured it. I checked the packages of 21 distro releases, started eleven of their resolvers from scratch with the default configuration (one inconclusive), tried 11 popular Docker images, the dnsmasq of 5 OpenWrt releases and 11 public resolvers, and left 10 resolvers running to watch the change live. What they did is, hour by hour, in What happened on 11 October, measured.
What DNSSEC is, on a real query
DNS without DNSSEC trusts whoever answers: if someone along the way changes the answer, your machine has no way to tell. DNSSEC adds signatures. Each zone signs its records with its keys, and the zone above publishes a digest of those keys (a DS record) signed with its own. That builds a chain from the root down to the name you asked for.
The one checking the chain is the validating resolver. To start, it needs to trust something without proof: the trust anchor, a copy of the key that signs the root. If the whole chain matches, the answer comes back with the AD flag (authenticated data). If anything does not match, it comes back SERVFAIL.
This is the real chain for gob.cl, asked directly to each server on 7 October:
- your resolver and its anchor
- KSK-2017 (20326), signs today
- KSK-2024 (38696), signs from 11 October
- zones below the root
- validated path
On Sunday the first link changed. The root’s change does not require anything below it to change: on 7 October .cl was signing with its key 21199 and gob.cl with 6334.
What changed on Sunday
Until 10 October the root signs its key set with KSK-2017, key tag 20326. From the 11th it is signed by KSK-2024, key tag 38696, which has been published in the root zone since January 2025. On 7 October the set had four keys (two zone signing keys, 8763 and 57780, because of a quarterly change in progress, plus the two KSKs) and a single signature, from 20326.
That set is cached for up to 48 hours, its TTL. That is why ICANN gives no failure time: a resolver that only trusts 20326 keeps working with its old copy and fails the next time it renews it. ICANN expects that to happen within the following 48 hours, although a lightly used resolver can take longer. According to Cloudflare, ICANN plans to revoke and remove KSK-2017 in 2027; ICANN’s document estimates the next change for 2029, and with it the move to ECDSA.
To see the mechanism I set up four unbound 1.26.1 resolvers that only differ in their anchor: one with the old key, one with the new one, one with both and one that updates itself with RFC 5011 and started on 7 October. The animation walks the six hops of the gob.cl chain with each one’s measured time (the sum of the six medians, asked directly to each authority from Chile, is 383 ms), and each resolver’s state is what the vigil measured on 7 October:
The first position is the state the vigil measured on 7 October (queries for example.com): the resolver that only has 38696 already fails. The second is what the standard predicts; the vigil keeps the four resolvers running through the change and this figure will be replaced with what it measures. The hop times are those of the gob.cl chain.
Two things the animation shows that are not obvious:
- Losing the root also breaks unsigned names. A resolver validating in strict mode has to prove from the root that the delegation of an unsigned name is insecure, and without the root it cannot. On 7 October the unbound that only trusts 38696 answered SERVFAIL for efraingaray.com too, which is not signed.
- RFC 5011 does not save what started late. The self-updating unbound saw 38696 on 7 October and put it on a 30-day hold (
ADDPEND). On Sunday it did not trust it yet, and afterwards it could not finish the hold on its own either: the set became signed only by a key it did not trust yet. With example.com, the vigil’s signed query, it started returning SERVFAIL about 7 h 38 min after the first round with the new key.
Who really validates
Most do not. APNIC measures with ads which users sit behind validating resolvers: in the 30-day window to 5 October, 9.1% in Chile (plus 3.6% validating partly) and 39.1% worldwide. If your resolver does not validate, the change does not touch you, although you may depend on one further up that does.
My own machines are a good example. On fedora (Fedora 43) and on the VPS (Ubuntu 24.04), systemd-resolved runs with DNSSEC=no: they do not validate, they delegate. The Mac uses 8.8.8.8. The VPS has Hostinger’s resolver configured, which does hand out validated answers. The router at fedora’s home returned the name with broken signatures with no SERVFAIL and no AD flag.
I asked the public resolvers from Chile, with the two RFC 8509 sentinel names and one name with deliberately broken signatures (two for Yandex). What you see from here is what the anycast node serving Chile answers:
- Cloudflare, Quad9 and AdGuard validate, and their sentinel answers like resolvers that already trust 38696.
- Google, OpenDNS, Control D, CleanBrowsing and DNS.SB validate, but answered both sentinel questions, so this test does not say which anchors they hold.
- Yandex DNS is the odd one. On its three addresses and in three repetitions (two of 81 queries got no answer), it sets AD on correctly signed data and lets two names with deliberately broken signatures through with NOERROR and no AD. Its sentinel answers like a resolver that trusts 20326 and not 38696, and the control question about 20326 comes out the other way round, as it should. That is consistent with validating in permissive mode: if that policy also applies when the root fails, on Sunday it could keep answering without setting AD instead of going down. The vigil will measure it.
Where DNSSEC is used
On the signing side, adoption is high at the top and low further down. I counted the whole root zone (serial 2026100702): 1,350 of 1,437 TLDs have a DS, .cl included. Among the top 1,000 of the Tranco list of 6 October (a popularity ranking that combines several sources), 109 have a DS record of their own. In a Chilean sample I picked by hand, 62 domains across six sectors, 7 are signed: gob.cl, nic.cl, santander.cl, bancointernacional.cl, falabella.com, sodimac.cl and wom.cl. None of the 8 universities or the 10 media outlets I picked.
- TLDs in the root zone93.9%1,350 of 1,437, .cl included
- Tranco top 1,00010.9%109 of 1,000 with their own DS
- Chilean sample11.3%7 of 62; 0 of 8 universities, 0 of 10 media
IANA root zone, serial 2026100702; Tranco list 26Y29 (6 October 2026); Chilean sample picked by hand, not random. A DS only counts if it is a DS record of the name itself.
Among TLDs, RSASHA256, algorithm 8, dominates: 1,086 publish it, against 269 publishing ECDSA P-256 (13), and 40 publish more than one. My domain is not signed either: efraingaray.com publishes no DNSKEY and .com returns a proof that there is no DS. For Sunday that does not matter: what counts is who validates.
Environment by environment
This is the measurement I was after. I separated two questions that usually get mixed: which key each package ships, and which key the resolver trusts when it starts with its default configuration.
| systemd-resolved | unbound-anchor | dns-root-data | unbound root.key | dnssec-root | BIND | dnsmasq | |
|---|---|---|---|---|---|---|---|
debian:10 | not found | 2010 and 2017 | 2017 | not found | not found | 2010 and 2017 | 2010 and 2017 |
debian:11 | 2017 | 2017 | 2017 and 2024 | not found | not found | 2017 | 2017 |
debian:12 | 2017 and 2024 | 2017 | 2017 and 2024 | not found | not found | 2017 and 2024 | 2017 |
debian:13 | 2017 and 2024 | 2017 and 2024 | 2017 and 2024 | not found | not found | 2017 and 2024 | 2017 and 2024 |
ubuntu:18.04 | 2010 and 2017 | 2010 and 2017 | 2010 and 2017 | not found | not found | 2010 and 2017 | not found |
ubuntu:20.04 | 2017 | 2010 and 2017 | 2017 and 2024 | not found | not found | 2017 | 2017 |
ubuntu:22.04 | 2017 | 2017 | 2017 and 2024 | not found | not found | 2017 and 2024 | 2017 and 2024 |
ubuntu:24.04 | 2017 | 2017 | 2017 and 2024 | not found | not found | 2017 and 2024 | 2017 and 2024 |
ubuntu:26.04 | 2017 and 2024 | 2017 and 2024 | 2017 and 2024 | not found | not found | 2017 and 2024 | 2017 and 2024 |
fedora:42 | 2017 and 2024 | 2017 and 2024 | not found | 2017 | not found | 2017 and 2024 | 2017 |
fedora:43 | 2017 and 2024 | 2017 and 2024 | 2017 and 2024 | 2017 and 2024 | not found | 2017 and 2024 | 2017 and 2024 |
fedora:44 | 2017 and 2024 | 2017 and 2024 | 2017 and 2024 | 2017 and 2024 | not found | 2017 and 2024 | 2017 and 2024 |
rockylinux:8 | 2010 and 2017 | 2017 and 2024 | not found | not found | not found | 2017 and 2024 | not found |
rockylinux:9 | 2017 | 2017 and 2024 | not found | 2017 and 2024 | not found | 2017 and 2024 | 2017 and 2024 |
rockylinux:10 | 2017 and 2024 | 2017 and 2024 | not found | 2017 and 2024 | not found | 2017 and 2024 | 2017 and 2024 |
amazonlinux:2 | none | 2017 | not found | 2017 | not found | 2010 and 2017 | not found |
amazonlinux:2023 | 2017 | 2017 | not found | not found | not found | 2017 and 2024 | not found |
alpine:3.18 | not found | 2017 | not found | not found | 2017 | 2017 and 2024 | 2017 |
alpine:3.20 | not found | 2017 | not found | not found | 2017 | 2017 and 2024 | 2017 |
alpine:3.22 | not found | 2017 and 2024 | not found | not found | 2017 and 2024 | 2017 and 2024 | 2017 and 2024 |
alpine:3.24 | not found | 2017 and 2024 | not found | not found | 2017 and 2024 | 2017 and 2024 | 2017 and 2024 |
Keys found in each package as offered on 7 October 2026, installed in the release's official image, by the year of the key (2010 = 19036, 2017 = 20326, 2024 = 38696). "Not found": the scan did not find that component in that release. This is what the package ships, not what a running resolver trusts: that is the next view. In the 13 images where its resolved.conf could be read, the default it documents for systemd-resolved is DNSSEC=no.
| Resolver and release | Version | Trusts 38696 right after starting? |
|---|---|---|
debian:11 bind9 | 9.16.50-1~deb11u6 | yes |
ubuntu:20.04 bind9 | 9.18.30-0ubuntu0.20.04.2 | yes |
ubuntu:24.04 bind9 | 9.18.39-0ubuntu0.24.04.7 | yes |
debian:12 unbound | 1.17.1-2+deb12u4 | yes |
ubuntu:24.04 unbound | 1.19.2-1ubuntu3.10 | yes |
amazonlinux:2023 unbound | 1.17.1-1.amzn2023.0.15 | no |
alpine:3.20 unbound | 1.20.0-r2 | no |
ubuntu:24.04 systemd-resolved DNSSEC=yes | 255.4-1ubuntu8.17 | cannot tell |
alpine:3.22 unbound | 1.25.2-r1 | yes |
fedora:42 unbound | 1.24.2-1.fc42 | no |
Default configuration, first start, asked with the RFC 8509 sentinel. In these tests BIND trusted every key in the root set signed by its anchor when it initialised its empty key store; unbound with RFC 5011 puts a key it has just seen on a 30-day hold, and with a fixed anchor it only trusts what that file has. systemd-resolved does not implement the sentinel.
| Image | Built | Trusts 38696? |
|---|---|---|
mvance/unbound:1.17.1 | 2023-10-05 | no |
mvance/unbound:1.20.0 | 2024-06-08 | no |
mvance/unbound:latest | 2024-10-19 | yes |
klutchell/unbound:1.17.1 | 2023-01-13 | no |
klutchell/unbound:latest | 2026-09-18 | yes |
isc/bind9:9.18 | 2026-06-17 | yes |
isc/bind9:9.20 | 2026-09-30 | yes |
cznic/knot-resolver:v5.7.4 | 2024-07-23 | yes |
cznic/knot-resolver:v5.7.8 | 2026-10-06 | yes |
pihole/pihole:2024.07.0 | 2024-07-05 | config: 2017 trust not measured: dnsmasq has no sentinel |
pihole/pihole:latest | 2026-09-19 | config: 2017 and 2024 trust not measured: dnsmasq has no sentinel |
| OpenWrt: what the dnsmasq-full package ships (contents, not a running test) | ||
OpenWrt 19.07.10 | 2.80-16.3 | 2010 and 2017 |
OpenWrt 21.02.7 | 2.85-9 | 2017 |
OpenWrt 22.03.7 | 2.86-16 | 2017 |
OpenWrt 23.05.6 | 2.90-2 | 2017 |
OpenWrt 24.10.8 | 2.93-r1 | 2017 and 2024 |
Images started with their default configuration (BIND forced to IPv4 and Knot non-interactive because of Docker; Pi-hole with DNSSEC turned on). Pi-hole and dnsmasq do not implement the sentinel, so for them the active trust-anchor lines were read from the generated configuration.
| Resolver | Validates? | Trusts 38696? |
|---|---|---|
Cloudflare 1.1.1.1 | yes | yes |
Google 8.8.8.8 | yes | cannot tell |
Quad9 9.9.9.9 | yes | yes |
OpenDNS 208.67.222.222 | yes | cannot tell |
AdGuard 94.140.14.14 | yes | yes |
Control D 76.76.2.0 | yes | cannot tell |
CleanBrowsing 185.228.168.9 | yes | cannot tell |
Yandex 77.88.8.8 | permissive | no |
DNS.SB 185.222.222.222 | yes | cannot tell |
Quad101 101.101.101.101 | no answer | · |
Lumen 4.2.2.2 | no | · |
Asked from Chile, one anycast node. "Cannot tell": the resolver validates and answered both sentinel questions, so this test does not say which anchors it holds. Permissive: it marks good answers with AD but let broken signatures through in this test.
What comes out of the four views:
- A package shipping the key does not mean the resolver trusts it. What counts is the file the configuration points to. A freshly installed unbound does not trust 38696 on Amazon Linux 2023 (it starts from its built-in anchor, with only the old key), on Fedora 42 (its root.key ships only 20326) or on Alpine 3.20, where the default configuration uses a fixed anchor, never updated: the trusted-key.key from the dnssec-root package (version 20190225), which only has the old key. On Alpine 3.22 it does.
- Old releases even ship the 2010 key. Debian 10, Ubuntu 18.04 and Amazon Linux 2 keep 19036, retired in 2019, and none ships 38696.
- On Debian 12 and Ubuntu 24.04, unbound comes out fine thanks to a service step. Before starting, unbound copies the root.key from dns-root-data, which ships both keys even though its built-in anchor is old.
- BIND learns at startup. Debian 11 ships BIND 9.16.50 with only 20326 in its files, and still, freshly started,
rndc managed-keys statusshows both keys as trusted from the start: when it initialises its empty key store it validates the DNSKEY set with the key it ships and installs the current ones. A key that appears later does wait 30 days. - Static anchors do not update themselves. dnsmasq does not do RFC 5011: it trusts the
trust-anchor=lines in its configuration. In Pi-hole v5.18.3’s configuration with DNSSEC, the active trust-anchor lines I found only have 20326; v6.4.3’s have both. OpenWrt’s dnsmasq-full package in 21.02, 22.03 and 23.05 ships only the old key, and 19.07’s also ships the 2010 one, retired in 2019; 24.10 ships both. - systemd-resolved validates only if you ask it to, and then does not update. In the 13 images where I could read its
resolved.conf, the default it documents isDNSSEC=no. WithDNSSEC=yesit uses its built-in anchor. 38696 reached systemd’s code in commit 8113361, released in version 258. The packages of Debian 12 and 13 and Fedora 42 already carry it in older versions; those of Ubuntu 22.04 and 24.04, Rocky 9 and Amazon Linux 2023 do not. - Three unbound images from 2023 and June 2024 start without trusting the new key: mvance/unbound 1.17.1 and 1.20.0 and klutchell/unbound 1.17.1. Their current versions do.
There are also repair paths I did not measure after the change, which I leave with their source. According to its manual, unbound-anchor (the service that runs before unbound on Fedora, RHEL and Amazon Linux) can recover the anchor by downloading root-anchors.xml from IANA over HTTPS and verifying it, if it runs and the check passes. On Debian and Ubuntu the service copies dns-root-data again at startup, which helps if that package is up to date. A fixed anchor needs someone to change it.
The vigil: ten resolvers crossing the change
Since 7 October ten resolvers have been running on my Mac: the four unbound from the animation, mvance/unbound 1.17.1 and its current version, Pi-hole v5 and v6 with DNSSEC on, and systemd-resolved from Ubuntu 24.04 and 26.04 with DNSSEC=yes. Every five minutes a script asks them a signed name, an unsigned one and both sentinel questions, and records which key signs the root according to a.root-servers.net. The same round adds Cloudflare, Google, Quad9 and Yandex.
On 7 October all of them answered except the one that only trusts the new key. Before the change I wrote here what I expected to go down: the old-key one, the late-started RFC 5011 one, mvance/unbound 1.17.1, Pi-hole v5 and Ubuntu 24.04’s systemd-resolved.
What happened on 11 October, measured
The root switched signatures between two rounds of the vigil: in the 13:59:24 UTC round the key set was still signed by 20326, and in the 14:04:29 UTC round only 38696 signed it. From that first round with the new key to the 19:21 UTC round on 12 October there are 338 rounds, counting that first one. A round starts roughly every five minutes and asks the resolvers one after another.
The table measures the vigil’s signed query, example.com. Times run from the first round with 38696 to the first round in which each resolver answered SERVFAIL:
| Resolver | First round with SERVFAIL | Time since the first round with 38696 | Up to the cut-off |
|---|---|---|---|
| systemd-resolved, Ubuntu 24.04 | 11 Oct, 14:24 UTC | about 20 min | 332 SERVFAIL out of 334 rounds; one stray validated answer at 14:35 UTC and one unanswered |
| unbound with RFC 5011, started 7 Oct | 11 Oct, 21:42 UTC | about 7 h 38 min | SERVFAIL in all of them |
| unbound trusting only 20326 | 12 Oct, 00:28 UTC | about 10 h 24 min | SERVFAIL in all of them |
| mvance/unbound 1.17.1 | 12 Oct, 03:33 UTC | about 13 h 29 min | SERVFAIL in all of them |
| Pi-hole v5 (2024.07.0) with DNSSEC | no SERVFAIL | 336 validated answers and 2 unanswered |
With the unsigned name, efraingaray.com, the times differ: mvance/unbound 1.17.1 had a stray SERVFAIL at 15:00 UTC on the 11th, answered fine for hours and has been SERVFAIL since 02:56 UTC on the 12th; the one trusting only 20326 started at 23:50 UTC on the 11th, the RFC 5011 one at 21:20 UTC and systemd-resolved at 14:55 UTC.
The unbound that only trusts the new key, the one that trusts both, current mvance/unbound and Pi-hole v6 kept validating without a single failure. Ubuntu 26.04’s systemd-resolved had one unanswered round and the rest validated. Cloudflare and Google answered validated every time there was an answer.
What those numbers say:
- Four of my five predictions held, and one did not, up to the cut-off. Pi-hole v5 keeps answering validated even though the only
trust-anchorline in its configuration is the one for 20326 and its upstream is 8.8.8.8 (captured from the container on 12 October). I do not know why it is still up. My hypothesis is that dnsmasq still uses the key set it validated before the switch, whose TTL is 48 hours; if so, it should go down around 14:00 UTC on 13 October, although the TTL does not guarantee it. I am keeping the vigil running past that time and will update this table. - Yandex stopped marking its answers as validated. Before the switch it marked 1,021 of 1,021 answers as validated. From 15:15 UTC on the 11th it started to alternate, and since 17:04 UTC none of 303 answers is marked as validated, although it still answers NOERROR. It fits what I measured on 7 October: its sentinel answered like a resolver that trusts 20326 and not 38696, and on broken signatures it lets the answer through instead of returning SERVFAIL. I cannot see its configuration, so this is a reading of its behaviour, not a certainty.
- No local resolver failed in the first round with the new key. They went down at very different times: about 20 minutes for systemd-resolved, between about 7 h 38 min and 13 h 29 min for the unbounds. None of them waited for the 48-hour TTL. The most likely explanation, to me, is that each one failed when it renewed its copy of the key set, but I did not log those renewals. What the table does show is that a check made an hour after the switch does not prove a resolver is fine.
- RFC 5011 did not rescue the late starter, as expected. It saw the new key on 7 October, put it on its 30-day hold and went down.
- Up to the cut-off, none of the four recovered in a sustained way. In the last round all four answer SERVFAIL also for efraingaray.com, which is not signed.
Quad9 left 100 of the 338 rounds after the switch unanswered, against 100 of 1,025 before (29.6% against 9.8%). Before and after, the gaps cluster mostly between 15:00 and 00:00 Chile time (79 and 86 out of 100), so I do not attribute them to the change, although I cannot rule out that it played a part. The two times three resolvers went unanswered in the same round after the switch, one was a systemd-resolved that asks Cloudflare, together with Cloudflare and Quad9: they share my network and partly the same upstream, so I do not know where the outage was.
How to check yours (and how to fix it)
First, ask your resolver the sentinel:
dig @YOUR_RESOLVER root-key-sentinel-is-ta-38696.dnstest.dev A
dig @YOUR_RESOLVER root-key-sentinel-not-ta-38696.dnstest.dev A
If the first answers and the second gives SERVFAIL, it already trusts 38696. If it is the other way round, it does not. If both answer, the test is inconclusive: it may not validate, may not implement the sentinel (dnsmasq, Pi-hole and systemd-resolved do not) or each question may reach a different server. For what you run yourself, check its anchors:
- unbound: if it uses
auto-trust-anchor-file(unbound-checkconf -o auto-trust-anchor-filetells you), look for 38696 in that file and check that it saysVALID, notADDPEND. If it uses a fixed anchor (trust-anchor-fileortrust-anchor:), that file or line must contain 38696. - BIND:
rndc managed-keys statusmust showkeyid: 38696withtrusted since. - dnsmasq and Pi-hole v5:
grep -r '^trust-anchor=' /etc/dnsmasq.d /etc/dnsmasq.conf. - systemd-resolved with DNSSEC=yes: if there are no
*.positivefiles in/etc/dnssec-trust-anchors.d/,/run/dnssec-trust-anchors.d/,/usr/local/lib/dnssec-trust-anchors.d/or/usr/lib/dnssec-trust-anchors.d/, it uses its built-in anchor.
To fix it, add the new key without removing the old one until the change. A key tag is just a 16-bit number, so compare the full DS record with the one IANA publishes. KSK-2024’s is this one:
. IN DS 38696 8 2 683D2D0ACB8C9B712A1948B27F741219298D0A450D612C483AF444A4C0FB2B16
In dnsmasq and Pi-hole v5 it goes as trust-anchor=.,38696,8,2,683D2D0A... with the full digest, or you upgrade to Pi-hole v6. In systemd-resolved it goes in a file /etc/dnssec-trust-anchors.d/root.positive with both DS lines, 20326 and 38696: according to its manual, as soon as you define an anchor for the root it stops using the built-in one, and it does not update by itself. In unbound, update the package that ships the anchor (dns-root-data on Debian and Ubuntu, dnssec-root on Alpine) or let unbound-anchor renew it. In BIND, updating the package does not change a key store that already exists: if rndc managed-keys status does not show 38696 as trusted, follow ISC’s guide for that case.
If something went down, ICANN’s emergency exit is to briefly turn validation off or set a negative trust anchor for the root (RFC 7646), install KSK-2024 and validate again.
What I cannot claim
- What happens outside my ten resolvers. The per-environment measurements are from 7 October, and what I saw on the 11th is the vigil: ten installs started from scratch, not the ones that have been running for years.
- The global state of public resolvers. I measured from a single place, Santiago, and I see the anycast node that serves me.
- Which anchors the resolvers without the sentinel hold. Google, OpenDNS, Control D, CleanBrowsing, DNS.SB and Hostinger’s resolver validate, but this test does not show what they trust.
- Installations that have been running for a while. I measured fresh starts. A BIND or an unbound with working RFC 5011, saved state and the 30-day hold completed already trusts 38696, even if its original package only shipped the old key. The risk is in losing that state (containers without a volume, reinstalls, recreated old images), in old fixed anchors and in updates that fail.
- A freshly started Fedora 43. Its root.key has both keys in
VALIDstate, but inside Docker its unbound answered SERVFAIL even for example.com, so that test does not count.
My take
For most people, who sit behind resolvers that do not validate or already trusted the new key, the change should have gone unnoticed. That was the expected outcome with a key published 21 months ahead, although I did not measure it. But the people who validate DNSSEC on their own are exactly the ones who set up a Pi-hole, an OpenWrt or an unbound in a container, and that is where I found old anchors in releases that are still around. What surprised me most was how differently each piece of software behaves with the same key: in these tests BIND trusted the new key from its first start, unbound with RFC 5011 put it on a 30-day hold and dnsmasq does not update it by itself.
When to worry and when not
- There is nothing to do on your machine if you use your ISP’s DNS, your stock router’s or a public one, or if your systemd-resolved runs with
DNSSEC=no: validation, if any, happens in the resolver above. - Check today if you have Pi-hole v5 with DNSSEC, OpenWrt up to 23.05 with dnsmasq-full and DNSSEC, unbound or BIND in containers that get recreated, Alpine 3.20 or older with unbound, or systemd-resolved with
DNSSEC=yeson Ubuntu 22.04 or 24.04. - Keep the emergency exit at hand if you run a resolver for other people: RFC 7646’s negative trust anchor or briefly turning validation off, then installing KSK-2024.
Measured on 7 October 2026 from Santiago, Chile. Containers on Docker 29.3.1 on an x86_64 Mac; test resolvers unbound 1.26.1-0+deb13u1 on debian:13-slim. Each distro’s packages installed that day from its repositories (Debian 10 and 11 from archive.debian.org). Tranco list 26Y29; root zone serial 2026100702; APNIC data with a 30-day window to 5 October.
Frequently asked questions
What changes on 11 October 2026?
The DNS root starts signing its key set with a new key, KSK-2024 (key tag 38696), instead of KSK-2017 (key tag 20326), which has signed since 2018. The new key has been published in the root zone since January 2025. A resolver that validates DNSSEC and only trusts the old key can no longer validate the root once it renews the key set, and from then on answers SERVFAIL, unsigned names included.
What happened on 11 October?
The root switched signatures between 13:59 and 14:04 UTC, according to two rounds of my vigil. For example.com, counting from the first round with the new key, Ubuntu 24.04's systemd-resolved with DNSSEC on started returning SERVFAIL about 20 minutes later; an unbound with RFC 5011 started on 7 October, after 7 h 38 min; one that only trusted the old key, after 10 h 24 min; and mvance/unbound 1.17.1, after 13 h 29 min. Up to 12 October none had recovered in a sustained way. The ones trusting the new key kept working, Pi-hole v5, whose configuration only carries the old key, was still validating the next day, and Yandex stopped marking its answers as validated.
Does it affect me if I use my ISP's DNS, Cloudflare or Google?
There is nothing to do on your machine: that resolver does the validation, and if it were not ready, fixing it falls to whoever runs it. I measured from Chile that Cloudflare, Quad9 and AdGuard already trust the new key; Google, OpenDNS, Control D, CleanBrowsing and DNS.SB validate, but this test does not show which anchors they hold. According to APNIC, only 9.1% of users in Chile sit behind a validating resolver.
How do I know whether my resolver already trusts the new key?
Ask it two RFC 8509 test names: dig @your-resolver root-key-sentinel-is-ta-38696.dnstest.dev and dig @your-resolver root-key-sentinel-not-ta-38696.dnstest.dev. If the first answers and the second gives SERVFAIL, it trusts it. If it is the other way round, it does not. If both answer, the test is inconclusive: it may not validate, may not implement the test (as with dnsmasq, Pi-hole and systemd-resolved) or each question may be answered by a different server, and its anchors have to be checked by hand.
I run Pi-hole with DNSSEC on. What should I do?
If it is Pi-hole v5 (the 2024.07.0 image ships v5.18.3), the active trust-anchor lines in its configuration only have 20326, and dnsmasq does not update itself: upgrade to v6, which writes both, or add the trust-anchor line for 38696 without removing the old one. With DNSSEC off, Pi-hole does not validate and the change does not affect it directly; it depends on its upstream resolver being ready.
Why can unsigned names fail too?
Because a resolver validating in strict mode has to prove, from the root down, that the delegation of an unsigned name is insecure, and once it needs to validate the root again it cannot. I measured it: on 7 October an unbound that only trusts the new key answered SERVFAIL for efraingaray.com as well, which is not signed.
What do I do if my DNS went down with the change?
ICANN's emergency advice is to briefly turn validation off or configure a negative trust anchor for the root (RFC 7646), then install KSK-2024 as an anchor and validate again. What you must not do is keep only 38696 before the change: that fails too, as I measured on 7 October.
Does my website need to do anything?
No. The change affects those who validate; those who sign or publish have nothing to do. My own domain, efraingaray.com, is not even signed: it publishes no DNSKEY and .com returns a proof that there is no DS.
Sources
- The scripts and raw data of every measurement: the vigil, the distro matrix, the fresh starts, the Docker images, OpenWrt, the public resolvers, the signed-domain count, every claim with its evidence and a summary of the review rounds.
- What to Expect During the Root KSK Rollover, ICANN, 27 July 2026: the schedule, the 48-hour TTL, who fails and how to recover (sections 1 to 3.4).
- ICANN publishes guidance for the October 2026 rollover, press release of 11 August 2026.
- The root anchors IANA publishes: 19036, 20326 and 38696, with their validity dates and digests.
- The keys to the Internet change on October 11, 2026, Cloudflare: the change, the sentinel and the dnstest.dev test names.
- RFC 4033, RFC 4034 and RFC 4035: DNSSEC; RFC 4034 appendix B defines the key tag and RFC 4035 section 5, validation.
- RFC 5011: automated trust anchor updates and the 30-day hold (section 2.4.1).
- RFC 8509: the trust anchor sentinel I used to ask each resolver which key it trusts.
- RFC 7646: negative trust anchors, the emergency exit ICANN mentions.
- RFC 7958: the root-anchors.xml format unbound-anchor downloads.
- The 2024-2026 Root Zone KSK Rollover: Initial Observations, Duane Wessels, Verisign: adoption of the new key as seen from the root.
- DNSSEC validation by country, APNIC Labs, and how they measure it.
- The systemd commit that adds KSK-2024 to resolved (4 March 2025) and version 258, the first release that has it.
- unbound-anchor manual, NLnet Labs: downloading root-anchors.xml when RFC 5011 is not enough.
- dnssec-trust-anchors.d, the systemd-resolved manual: where it reads anchors from, that it stops using the built-in one as soon as you define one, and that it does not update them from DNS.
- Simon Kelley, dnsmasq’s author, on RFC 5011, dnsmasq-discuss list, 15 October 2018: dnsmasq does not implement it.
- Root KSK Rollover in BIND 9, ISC, and DNSSEC validation on BIND, SIDN, on the initial start of the key store.
- OpenWrt 25.12’s dnsmasq, version 2.93, the same as 24.10.8.
- I built a fully post-quantum handshake: the other cryptography measurement from Chile on this blog.
- The OSI model explained: where DNS lives in the stack.
NEWSLETTER · MEASURED
A new measurement, when there is one.
Each post with its numbers, its failures and the scripts to repeat it. One to three emails a month. Unsubscribe with one click.
LATEST
- Grid jolt every minutepeak 42.6× lower at random
- SMS of DeathAndroid 12 to 17
- Unlimited-OCR 2 on 16 GBtie with Chandra 2, out of memory at 80 pages
I run Listmonk on my own server; your address stays there. You confirm by email before getting anything.
Comments
No comments yet. The first one is yours.