CVE-2026-47703: AdGuard Home: DoQ-to-UDP State Reduction and Source-Port Oracle
Summary This report covers the client-triggered DoQ forwarding path in: - dnsproxy v0.81.2 (adguard/dnsproxy:v0.81.2) - AdGuard Home v0.107.74 (adguard/adguardhome:latest, image version label v0.107.74) The issue was reproduced on 2026-04-25 with the products configured through their documented DoQ listener and plain UDP upstream surfaces.
Published Jun 4, 2026 · Updated Jul 20, 2026
Affected packages
github.com/AdguardTeam/AdGuardHome
- Ecosystem
- Go
- Affected
- <= 0.107.74
- Fixed in
- 0.107.75
- Check
go list -m github.com/AdguardTeam/AdGuardHome
github.com/AdguardTeam/dnsproxy
- Ecosystem
- Go
- Affected
- < 0.81.3
- Fixed in
- 0.81.3
- Check
go list -m github.com/AdguardTeam/dnsproxy
Advisory details
From the GitHub Advisory Database (CC BY 4.0).
Summary
This report covers the client-triggered DoQ forwarding path in:
dnsproxyv0.81.2(adguard/dnsproxy:v0.81.2)AdGuard Homev0.107.74(adguard/adguardhome:latest, image version labelv0.107.74)
The issue was reproduced on 2026-04-25 with the products configured through their documented DoQ listener and plain UDP upstream surfaces. The scope is the internal backend UDP hop created when a DoQ query is forwarded to a udp:// upstream.
On that path, the backend DNS ID is not preserved as an independent source of entropy. For both products, the backend observer saw dns_id=0 for every sampled client-triggered query on the tested path. Repeated reruns then showed the same txid=0 behavior and the same positive source-port oracle on every sampled run. A separate quoted-port ICMP oracle distinguished the correct backend UDP source port from a wrong one with a stable, client-visible behavior change.
Attached evidence:
dnsproxyoracle path onv0.81.2: [attachments/artifacts/g03/20260425T141500Z-g03-v0812/summary.txt](attachments/artifacts/g03/20260425T141500Z-g03-v0812/summary.txt)dnsproxyv0.81.2repeatability: [attachments/artifacts/g03/repeatability-v0812.md](attachments/artifacts/g03/repeatability-v0812.md)dnsproxysteering follow-up onv0.81.2: [attachments/artifacts/g04/20260425T141900Z-g04-v0812/summary.txt](attachments/artifacts/g04/20260425T141900Z-g04-v0812/summary.txt)AdGuard Homeoracle path: [attachments/artifacts/g05/20260425T113000Z-g05/summary.txt](attachments/artifacts/g05/20260425T113000Z-g05/summary.txt)
Root Cause Analysis
The observable behavior is consistent across both products:
- A DoQ client query is accepted on the frontend listener.
- The query is forwarded over a backend UDP leg.
- On that backend leg, the forwarded DNS
IDcollapses to0on the client-triggered path instead of remaining a fresh per-query variable. - The backend UDP source port is still allocated per query.
- When an ICMP error quotes the actual backend source port, the forwarding path flips behavior in a way that does not occur for a wrong quoted port.
That combination removes txid from the backend tuple on the tested path and leaves the UDP source port as the main remaining variable. In practical terms, the backend hop stops behaving like a fresh (txid, source-port) pair per forwarded query and instead becomes a one-variable state exposure.
For dnsproxy, the correct quoted port does more than produce a failure signal: it can push resolution away from the primary UDP upstream and into the fallback upstream. For AdGuard Home, the same condition produces a fast SERVFAIL.
Reproduce
Prerequisites:
- Docker and Docker Compose
- OpenSSL
- build the lab helper image used by the attached harness and observer
The attached reproducer bundle contains only the files needed for this report:
- scripts:
attachments/scripts/ - helper image build files:
attachments/docker/unbound-doq-attacker/ - compose files:
attachments/docker-compose.g03.yml,attachments/docker-compose.g04.yml,attachments/docker-compose.g05.yml - shipped evidence:
attachments/artifacts/...
Build the helper image first:
cd attachmentsdocker build -t unbound-doq-attacker:latest -f docker/unbound-doq-attacker/Dockerfile docker/unbound-doq-attacker
To rerun dnsproxy:
cd attachmentsbash scripts/repro-g03-dnsproxy-oracle.sh- Inspect
artifacts/g03/<RUN_ID>/summary.txt - Inspect
artifacts/g03/<RUN_ID>/entropy-backend.jsonl,txid_correct-backend.jsonl, andport_correct-backend.jsonl
To rerun the dnsproxy fallback-steering case:
cd attachmentsbash scripts/repro-g04-dnsproxy-steering.sh- Inspect
artifacts/g04/<RUN_ID>/summary.txt - Inspect
steering_correct-main.jsonlandsteering_correct-fallback.jsonl
To rerun AdGuard Home:
cd attachmentsbash scripts/repro-g05-adguardhome-oracle.sh- Inspect
artifacts/g05/<RUN_ID>/summary.txt - Inspect
entropy-backend.jsonl,txid_correct-backend.jsonl, andport_correct-backend.jsonl
The attached evidence includes fresh dnsproxy v0.81.2 reruns, one official- profile AdGuard Home run, and the minimal reproducer bundle used by both.
Impact
For both products, the tested DoQ-to-UDP path is no longer a full (txid, source-port) search surface:
dnsproxy: four of four sampled runs showedtxid=0on the backend hop and a positive source-port oracle onv0.81.2. The remaining unknown isport_only. Median wrong/correct port latency was327.99 ms / 40.93 ms.AdGuard Home: four of four sampled runs showedtxid=0on the backend hop and a positive source-port oracle. The aggregate again classifies the remaining unknown asport_only. Median wrong/correct port latency was319.14 ms / 37.02 ms.
Product-specific effects:
dnsproxy: a correct port guess produced an empty client-visible answer on the base oracle path, and in the fallback profile it steered all eight tested queries away from the main upstream and into the fallback upstream.AdGuard Home: a correct port guess produced fastSERVFAILand an extra backend query.
This is the security-relevant point. On the tested official profiles, the backend hop no longer forces an off-path attacker to deal with two fresh random fields per forwarded DNS race. The DNS ID is already known: it is deterministically 0 on the client-triggered DoQ-to-UDP path. The only remaining backend tuple variable is the UDP source port, and the attached evidence shows a repeatable oracle for that remaining variable.
That places the path in the same threat-model class as oracle-assisted DNS forgery work such as SAD DNS and TUdoor: the attack first uses an oracle to learn or validate the tuple state that protects an off-path response race, and only then attempts the forged response. This report stops short of a forgery demo, but the evidence already shows the crucial precondition on the tested backend hop: the tuple is not high-entropy anymore. It has been reduced from (txid, source-port) to source-port only.
---
Attachments attachments.zip
Credited to N0zoM1z0.
References
- https://github.com/AdguardTeam/AdGuardHome/security/advisories/GHSA-xgx4-4h9w-53pv
- https://github.com/AdguardTeam/dnsproxy/commit/f00d992ce9567a50f596853978ad6500acfdcf1d
- https://nvd.nist.gov/vuln/detail/CVE-2026-47703
- https://pkg.go.dev/vuln/GO-2026-5756
- https://github.com/advisories/GHSA-xgx4-4h9w-53pv
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. No analysis has been written for this entry yet. Methodology
