I probed github.com's SSH service twice in one week, with the same tool and the same options. The first run, from my Ubuntu 24.04 laptop, said two post-quantum key exchanges worked only with the experimental Open Quantum Safe client. The second run, from a Debian 13 container, said the stock client and the experimental client negotiate identically.
GitHub had not changed anything. My client had. Debian 13 ships a newer OpenSSH, and that release already does what I was using an experimental build for.
This post is about that shift: stock OpenSSL and OpenSSH now negotiate post-quantum key exchange out of the box, which changes what a "stock versus experimental" comparison can tell you, and what a post-quantum audit should actually measure.
Part 4 of a series on building pqc-probe. Earlier: part 1, part 2, part 3.
What changed in the stock clients
Post-quantum key exchange arrived in the mainstream tools in a few releases:
| Date | Release | Post-quantum key exchange |
|---|---|---|
| Apr 2025 | OpenSSH 10.0 | mlkem768x25519-sha256 becomes the default |
| 8 Apr 2025 | OpenSSL 3.5.0 | default TLS key shares become X25519MLKEM768 and X25519 |
| 19 Sep 2024 | OpenSSH 9.9 | adds mlkem768x25519-sha256, enabled by default; adds the name sntrup761x25519-sha512 |
| Apr 2022 | OpenSSH 9.0 | sntrup761x25519-sha512@openssh.com becomes the default |
The OpenSSL change is easy to miss because it is one line in the source. The 3.5 default group list begins with ?*X25519MLKEM768, and the * means the client sends an ML-KEM hybrid key share in every handshake, next to a classical X25519 one. A server that understands ML-KEM can complete a post-quantum handshake without a single extra round trip. OpenSSL 3.0, still common on long-term-support distributions, has no ML-KEM at all.
OpenSSH went further in 10.1, which by default warns the user when a connection is not using a post-quantum key exchange.
Same server, two client machines
GitHub's SSH service offers three post-quantum key exchanges: mlkem768x25519-sha256, sntrup761x25519-sha512 and its older name sntrup761x25519-sha512@openssh.com. Here is what pqc-probe reported against it from each machine:
| Ubuntu 24.04 laptop | Debian 13 container | |
|---|---|---|
| Stock SSH client | OpenSSH 9.6p1 | OpenSSH 10.0p2 |
mlkem768x25519-sha256 with the stock client | not implemented | completed |
sntrup761x25519-sha512 with the stock client | not implemented under this name | completed |
sntrup761x25519-sha512@openssh.com with the stock client | completed | completed |
| Stock versus OQS client diff | 2 algorithms "only experimental" | none |
The laptop's OpenSSH 9.6 predates 9.9, so it lacks ML-KEM and the newer sntrup name. It still completes a post-quantum exchange through the older @openssh.com name, which every OpenSSH since 9.0 uses by default. The Debian container's 10.0 does all three.
The report from the laptop was correct; it described the laptop. Read naively, though, it suggested that GitHub needs an experimental client for ML-KEM, which is not true for anyone on a current OpenSSH.
What the diff still tells you
A "stock versus experimental" diff measures the gap between two clients against one server. When the stock client was classical-only, that gap was a fair stand-in for "what the server supports beyond the mainstream". With post-quantum stock clients, it splits into two different findings, and pqc-probe now labels them apart:
- Newer mainline, not experimental. An algorithm only the OQS client negotiates may simply be one your stock client is too old for. pqc-probe marks these with the first mainline release that has them, for example
mlkem768x25519-sha256 (stock OpenSSH since 9.9). The fix is to update the client, not to adopt an experimental build. - Genuinely experimental. Names such as
frodokem-976-shake-sha384or the@openquantumsafe.orghybrids exist only in OQS builds. A server that offers them is running experimental software, which is worth knowing for a different reason.
For TLS the effect is stronger. A stock OpenSSL 3.5 client already completes X25519MLKEM768 wherever the server allows it, so an empty TLS diff no longer means the server is classical. When the stock client itself negotiated a post-quantum group, pqc-probe says so in the report (vanilla_pq_capable), because the diff then says little about the server.
What it means for an audit
A connection is post-quantum only when both ends support it. A server audit answers half the question; the clients that actually connect answer the other half.
- Measure the server with a current client. Use OpenSSL 3.5+ and OpenSSH 10+ (or pqc-probe's Docker image, which carries both). An old client makes a capable server look incapable.
- Know your client population. Long-term-support systems keep their versions for years: Ubuntu 24.04 ships OpenSSH 9.6, which has sntrup761 but no ML-KEM, and OpenSSL 3.0, which has no post-quantum key exchange at all. Build agents, deploy scripts, backup jobs and the SSH clients on your own servers count too. Until they are upgraded, a server's ML-KEM support does nothing for their TLS connections, and their SSH connections stay on the older sntrup761 exchange.
- Record the client with the result. A finding such as "ML-KEM not negotiated" is meaningless without the client version that produced it. pqc-probe's JSON reports include the client versions for exactly this reason.
The practical consequence: "our servers support post-quantum key exchange" and "our connections are post-quantum" are two different claims, and only the second one protects traffic today against harvest-now, decrypt-later collection.
The comparison is shrinking, and that is good news
When I started pqc-probe, the interesting output was the diff: what the experimental client could do that the stock one could not. For the most common algorithms, that gap has largely closed. ML-KEM hybrids are in stock OpenSSL and OpenSSH, on by default, and negotiated without anyone asking for them.
That moves the useful question from "can it be done?" to "is it being done, by every client that matters?" The answer now depends far more on how old your clients are than on anything exotic on the server, and the cheapest post-quantum upgrade available to most teams is a routine package update.
Next in the series: shipping two liboqs builds in one container, and two GCC 14 bugs found on the way.
