Is github.com/AdguardTeam/dnsproxy safe?

GHSA-xgx4-4h9w-53pv

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. 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 dnsid=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:

• dnsproxy oracle path on v0.81.2: [attachments/artifacts/g03/20260425T141500Z-g03-v0812/summary.txt](attachments/artifacts/g03/20260425T141500Z-g03-v0812/summary.txt)
• dnsproxy v0.81.2 repeatability: [attachments/artifacts/g03/repeatability-v0812.md](attachments/artifacts/g03/repeatability-v0812.md)
• dnsproxy steering follow-up on v0.81.2: [attachments/artifacts/g04/20260425T141900Z-g04-v0812/summary.txt](attachments/artifacts/g04/20260425T141900Z-g04-v0812/summary.txt)
• AdGuard Home oracle path: [attachments/artifacts/g05/20260425T113000Z-g05/summary.txt](attachments/artifacts/g05/20260425T113000Z-g05/summary.txt)

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 ID collapses to 0 on 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.

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 attachments
• docker build -t unbound-doq-attacker:latest -f docker/unbound-doq-attacker/Dockerfile docker/unbound-doq-attacker

To rerun dnsproxy:

• cd attachments
• bash scripts/repro-g03-dnsproxy-oracle.sh
• Inspect artifacts/g03//summary.txt
• Inspect artifacts/g03//entropy-backend.jsonl,

txidcorrect-backend.jsonl, and portcorrect-backend.jsonl

To rerun the dnsproxy fallback-steering case:

• cd attachments
• bash scripts/repro-g04-dnsproxy-steering.sh
• Inspect artifacts/g04//summary.txt
• Inspect steeringcorrect-main.jsonl and steeringcorrect-fallback.jsonl

To rerun AdGuard Home:

• cd attachments
• bash scripts/repro-g05-adguardhome-oracle.sh
• Inspect artifacts/g05//summary.txt
• Inspect entropy-backend.jsonl, txidcorrect-backend.jsonl, and

portcorrect-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.

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 showed txid=0 on the backend hop and

a positive source-port oracle on v0.81.2. The remaining unknown is portonly. Median wrong/correct port latency was 327.99 ms / 40.93 ms.

• AdGuard Home: four of four sampled runs showed txid=0 on the backend hop

and a positive source-port oracle. The aggregate again classifies the remaining unknown as portonly. Median wrong/correct port latency was 319.14 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 fast SERVFAIL and 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