Every post-quantum migration starts with the same question: does this server support a post-quantum key exchange yet? It sounds like a yes or no. It isn't, because the honest answer depends on which client is asking.
I built pqc-probe to answer it from real handshakes instead of configuration files or advertised lists. It connects to a server over TLS and SSH, tries each key exchange one at a time, and reports which ones actually complete. Version 0.1.0 is out this week as a CLI, a web service, a Docker image and a GitHub Action.
This post covers what the tool measures, how to read its output, and what it found across the 1,000 most popular sites. Later posts in the series go deeper into the surprises along the way.
Offered is not negotiated
A TLS client sends a list of key exchange groups it can use, and the server picks one. A normal connection tells you only which group won, not which others the server would have accepted. So pqc-probe offers one group at a time and records whether the handshake completes:
openssl s_client -connect github.com:443 -groups X25519MLKEM768 -4 If that handshake succeeds, the server supports X25519MLKEM768. If it fails, the server does not, whatever any configuration page claims. The probe repeats this for seven groups: four classical (X25519, P-256, P-384, P-521) and three ML-KEM hybrids (X25519MLKEM768, SecP256r1MLKEM768, SecP384r1MLKEM1024).
SSH needs a different trick. The server sends its full list of key exchange algorithms before any encryption starts, so pqc-probe reads that packet directly. Then, for each post-quantum algorithm both sides know, it runs a real key exchange and marks the row ok (tested) only when the exchange completes. Being on the server's list is not enough.
Two clients, one diff
Which client asks matters, so pqc-probe asks with two:
| Mode | TLS client | SSH client | Answers |
|---|---|---|---|
vanilla | stock OpenSSL 3.5 | stock OpenSSH | What a normal production client gets today |
experimental | OpenSSL + oqs-provider | OQS-OpenSSH | What the Open Quantum Safe algorithms unlock |
--mode both runs the two side by side and prints a diff. A group that only the experimental client can negotiate means the server supports something your users' clients cannot use yet. A group both can negotiate is deployable today.
There is one catch, and the tool flags it. OpenSSL 3.5 already supports ML-KEM hybrids out of the box. When the stock client is itself post-quantum capable, an empty diff says little about the server, so the report adds a note saying exactly that.
Reading a real report: github.com
Here is GitHub, probed from the Docker image with both protocols and both client stacks (trimmed):
$ docker run --rm --init ghcr.io/asifmahbubnaeem/pqc-probe:0.1.0 \ pqc-probe probe github.com --protocol both --mode both TLS KEMs (vanilla vs experimental): offered vanilla experimental X25519 ok ok P-256 ok ok P-384 - - P-521 - - X25519MLKEM768 - - SecP256r1MLKEM768 - - SecP384r1MLKEM1024 - - SSH KEX (server offer × client support): server offers mainline ssh OQS-OpenSSH mlkem768x25519-sha256 ok (tested) ok (tested) sntrup761x25519-sha512 ok (tested) ok (tested) sntrup761x25519-sha512@openssh.com ok (tested) ok (tested) curve25519-sha256 ok ok ... SSH diff: Mainline ssh and OQS-OpenSSH negotiate identically on this target. Two services on the same domain give opposite answers:
- TLS (port 443): only X25519 and P-256 complete. The stock OpenSSL 3.5.7 offered ML-KEM, and GitHub's front end declined it. At the time of this run, a browser connecting to github.com got a classical key exchange.
- SSH (port 22): three post-quantum key exchanges complete with both clients, so
git pushover SSH is already post-quantum.
The SSH diff is empty because the image's stock client is OpenSSH 10.0, which already supports these algorithms. The same probe from an older laptop with OpenSSH 9.6 showed two of them as "only experimental". That is the client changing, not the server.
What the top 1,000 sites do
Tranco list K9YYW, top 1,000, scanned 2026-11-06 with pqc-probe scan · one vantage point, IPv4 only · data and method in results/ in the repository]
Among the most popular sites that answer over TLS 1.3, post-quantum key exchange is already the majority. Read "accepts" strictly: the handshake completes when X25519MLKEM768 is offered alone, which shows support, not that the server prefers it.
The 305 sites without a handshake say nothing about post-quantum support either way: most are list entries with no DNS answer or no web server on port 443. The full dataset, the retry passes and the charts are in results/.
Three things that surprised me
- The SSH debug log cut the server's offer in half. My first version read the server's algorithm list from
ssh -vv. On a test server with a long list, the line simply stopped mid-name: OpenSSH truncates debug lines at about 1 KB. Reading the key exchange packet directly fixed it. - Testing too fast produced false negatives. OpenSSH 9.8 and later penalise addresses that open many short connections (
PerSourcePenalties). My test server's log showed a 16-second penalty kicking in, after which real support looked like failure. The probe now tests at most 8 algorithms, one per second. - The stock client is already post-quantum. OpenSSL 3.5 and OpenSSH 10 ship ML-KEM by default, so the classic "vanilla versus experimental" comparison is shrinking. Where it still shows a gap, the gap is usually an old client, not the server.
Each of these gets its own post in this series.
Try it
The Docker image carries both client stacks, so it is the quickest start:
docker run --rm --init ghcr.io/asifmahbubnaeem/pqc-probe:0.1.0 \ pqc-probe probe example.com --protocol both --mode both With Rust 1.89 or later, the CLI installs from crates.io and uses the openssl and ssh already on your machine:
cargo install pqc-probe pqc-probe probe example.com --protocol both Static Linux binaries are on the GitHub release page.
To check your own hosts on every deploy, the GitHub Action keeps a JSON report per host and can fail the job when a host has no post-quantum key exchange:
- uses: asifmahbubnaeem/pqc-probe@v0.1.0 with: targets: api.example.com, git.example.com:2222 require-pq-ssh: "true" Every report follows a published JSON schema, which only ever gains fields within a version, so you can build on it. Please probe hosts you are responsible for.
What's next
The rest of this series:
- Your SSH debug log is truncating the server's offer
- Probing SSH without getting penalised
- The stock client is already post-quantum
- Shipping two liboqs builds in one container, and two GCC 14 bugs found on the way (#210, #211)
- Running a scanner for strangers: SSRF, DNS rebinding and rate limits
pqc-probe is the first project of three. pqc-bench measures what post-quantum TLS costs in performance, and a migration dashboard built on pqc-probe's reports comes next.
The code is on GitHub under MIT or Apache-2.0. Issues and results from your own servers are very welcome.
