▲ 61 points
back
103 comments
What I do is laughably simple.
1. Disable all logging about break-in attempts.
2. Do not have any common user names like "root".
Say you want to be able to log in as root from anywhere, just with a password. This is a wise idea; what if you need access, but are in a situation where you are not able to use a certificate?
Make up an alternative name like roto-rooter or whatever pops into your head. Install it into the password file as an alternative name for UID 0. (Make sure it appears later than the "root" entry!). Also edit the shadow file, making sure that the entry is duplicated for the alternative name.
Then in the sshd config file, use AllowUsers to allow only a whitelisted set of users. Here if you say "AllowUsers rotorooter", then the only user id that can authenticate is exactly that one, and it's mapped to UID 0 via passwd/shadow. Add any other accounts you would like to remotely access, giving them similar aliases if they happen to land into a commonly probed space, or are something that a targeted attacker could infer from knowing something about you.
Attackers do not probe the user ID space at all. They concentrate exclusively on probing the password spaces of common user IDs like root, admin, database, www-data, etc. If your system does not support any of those IDs, they are not on a trajectory to crack anything. Your rotorooter password could be "g0d" and they will not get in, if all they ever try is root.
This is a bad idea™. You should never have more than one UID 0 on a Unix system. This will violate most corporate security guidelines (STIG and CIS) *. And for good reason.
A much better idea is to set up a non-root user and configure sudo correctly.
* https://www.stigviewer.com/stigs/red_hat_enterprise_linux_9/...
Port knocking, and other bespoke middle layers in front of internet services, is stupid. It violates Kerckhoffs’s principle¹. If you want more secret bits which users need to know in order to access your system, increase your password lengths, or cryptographic key sizes. If you want to keep log sizes (or “noise”) manageable, adjust your logging levels.
Anything added in front of your normal service also complicates access, since it’s non-standard. If you want a standard solution to solve all your needs for secure access of IP-based services, use IPsec and be done with it once and for all.
1. <https://en.wikipedia.org/w/index.php?title=Kerckhoffs%27s_pr...>
(Adapted from this old post: <https://news.ycombinator.com/item?id=39898061>)
Agreed.
And for logging in particular: Just switch the logging over to a temp file that lives in RAM (what disk thrash/write-amplification/SSD wear?), or even disable it altogether for failed login attempts.
We already know that there are great hordes of zombies outside of the castle, banging on the doors and the blocked-off spaces where the windows once were, picking away tirelessly. That's been a constant for many years. Documenting their continued persistence is pretty meaningless. None of it is actionable, or stoppable. It's just going to keep happening. Recording attempts from valid users is also largely without merit; it also just looks like noise, and we've got other ways to troubleshoot stuff that breaks without maintaining a long list of zombie attacks to peruse.
If a zombie actually manages to get in, then that's pretty important to keep track of; log that. But the attempts don't mean anything and have not meant anything for a very long time.
(When the word comes forth that the zombies are gone and the noise has ceased, it will be broadcast so far and wide that even the most noise-deafened sysadmins will find it impossible to ignore. In that seemingly-impossible unlikelihood, we can then resume recording attempts to log in with ssh.)
I agree port knocking is a direct violation of Kerckhoff's principle. However, the proposed solution has non-discoverability from unauthorized sources which isn't necessarily in the threat model of OpenSSH or general cryptography. I do feel like this is potentially a desirable trait. I elaborated a bit more here[0], but I'm curious if you have any grander thoughts on how this could be approached
It violates Kerckhoffs’s principle¹
Appeal to authority?
In the real world, it doesn't matter. Anything that makes the attacker's life harder is fair game. Stupid dogmatic sheep-like mindlessness only leads to "herd exploitability".
also complicates access
That's the whole point.
No, it doesn't. You're arguing against a straw man. The primary motivation was in second paragraph.
It is not stupid. Your stated principle assumes the exposed OpenSSH server does not have pre-auth 0-days. Having open 22 exposes a lot of information about the machine already, not to mention possible banners. The mentioned method is an elegant solution to both risks.
I think you are missing the point. It isn't about increasing the bits of security, it's about avoiding the flood of port scanners and the logs that produces.
IPsec is way more complicated to set up than this (or other VPN solutions like Wireguard or OpenVPN for that matter), and doesn't even completely solve that problem, because your ipsec port is open. Although, admittedly, there are probably less bots looking for ipsec than ssh.
A scheme like this is in the process of being standardised by Peter Gutmann who knows what he is doing. To give that statement more context he's an NZ cryptographer and has also worked in standardisation before. So, he is promoting this model with knowledge of cryptography, risks and the standards process.
Is this actually practically reducing the attack surface? We're replacing a battle hardened service with a random one that has the ability to manipulate the firewall rules.
No, it's not. The article itself is somewhat incoherent; for instance, if you're using key-only authentication (which: of course you are) fail2ban has literally no function (unless you think attackers are brute-forcing ECDH keys).
The reality is that there are basically two ways to operate SSH:
(1) You can, because OpenSSH is the significant remote service with the literal best track record of any remote service, just disable passwords and let SSH run in 22/tcp exposed to the Internet. Probably stop logging people scanning you; there's nothing you're going to do about it, so it's not real information.
(2) You can keep SSH behind WireGuard, an even simpler security protocol with an even better security story (though: OpenSSH is quite solid), which is designed to not to chat with counterparties that don't have keys, even to do negotiation.
Everything else is performative.
I'd incline towards option (2).
I think the answer is yes. A response from a remote system (or lack of) is a signal. The conventional ssh protocol method when connecting to a server is to instantly send over the identification string. RFC 4253 requires "both sides MUST send an identification string". It doesn't necessarily decide an order, but most ssh implementations I've seen send their protocol string before any other data has been received.
The broadcasting of the versioning information does constitute a potential leakage of information that could be useful to an attacker. Even if we contrast this to something like mTLS, the client certificate doesn't come until fairly late in the handshake, so there is still information that can be cleaned from the ServerHello from an unauthenticated inspector. This is also the case with QUIC since it piggybacks off the general TLS handshake.
I think the issue is, for a known set of systems, can you create communications between them that are oblivious/non-discoverable to non-authorized systems. I think the answer is yes, but it requires an out-of-band key agreement protocol. Wireguard is an example. However, the problem of out-of-band key agreement can't really be ignored.
I think the article's method is somewhat valid. I also think it is non-ideal for only doing source IP based rulesets, especially in the world of IPv4 and NAT being prevalent.
A sensible observation. In theory, the entire codebase for a simple HMAC knock knock ought to be tiny and easy to harden with multiple rounds of human and LLM review.
The goal should be to have your complex layers sitting in front of simpler, easier-to-review lines of defence. Once you get to something like an open connection to ssh, the potential attack surface would be orders of magnitude larger, even though it’s more mature and closely scrutinised.
If you mean fail2ban by “battle hardened service”, it manipulates the firewall rules as well. And OpenSSH has the same functionality built-in, without any additional service (PerSourcePenalties).
I don’t know why this wasn’t mentioned before but why not use a Firewall. If you’re using a virtual box like Hetzner or Scaleway you can specify an ip or range at the router level. For all intents and purposes this removes public exposure. Scaleway also has a cheap VPN bridge. So you never need to connect via the public internet if you don’t want to… hardly gets more secure than that
If you are connecting from a residence or a mobile device (like a laptop), your ip address isn't fixed, so a firewall won't help.
They could be hosting on their own homelab network (without tailscale etc magic)
I made my ssh server to listen ipv6-only, and it has been super silent in the logs ever since.
In the first iteration the IPv6 got polled by a handful of attackers as soon as the letsencrypt certificate was published. In the second iteration I just picked another IPv6 address from the /64 and made ssh.example.com to point to it. This should work until the attackers start guessing subdomain s...
This is an interesting piece that is often overlooked by folks in the "but NAT is security" camp; Having a sparse address space that is 64-bits makes it impossible to iteratively scan over a range. If you don't reverse resolve or you disallow zone transfers then you also have no real discoverability for that /64.
The real solution is using something like Pangolin or Tailscale (or Headscale) for this. You can control access way better and you never have to expose ssh ever. Not even temporarily.
I agree with you.
I’m reading this thread and wondering if I’m missing something, why people are still talking about port knocking, port obfuscation, and fail2ban.
I use a cloud VPS. I ssh in via Tailscale. The cloud provider firewall blocks all incoming connections except traffic originating from Cloudflare IP ranges on port 443. My host plays dead to portscans. I check with nmap periodically. I have a break-glass backup terminal login option via my cloud provider dashboard (secured with MFA) in case Tailscale failed and needed investigation and repair.
This has worked well for me on AWS and Oracle Cloud. It’s quick and easy to set up. You need a timed service to refresh the ingress IP ranges, but any LLM could spit that out in a second.
> If a zero-day drops in OpenSSH...
Realistically, fwknop is more likely to have a vuln than OpenSHH. Last release was two years ago and the readme dates back twelve :/ Time will tell.
Just wrap it in wireguard. WG won't respond to any incoming handshake without a valid and authorized client key, it's essentially invisible.
The enlightened trick is to run ssh on port 443.
I would go with something like:
seed=`date +%s%N`; ( echo "secretknock-20260811|$seed"|sha512sum ; echo $seed ) | xargs nc -u 192.168.1.1
On other hand if threat vector includes network monitor with ability to replay i would use wireguard to wrap ssh traffic.
> and why it is not enough on its own
Nothing is good enough on its own.
Geoblocking, fail2ban, port obscurity, SSH keys, limiting logins to specific usernames, not using your public internet nickname, putting things behind CloudFlare tunnels or WireGuard, wildcard DNS obscurity, 2FA... There are many options.
Defense in depth is the only way to put services on the internet.
Just changing the SSH port is probably enough for small VPS.
Port 22 is getting perma-DDOSed on public IPv4 addresses so use it at your own peril but at least for now most bots don't bother port scanning everyone.
Seems like a wireguard port in would be a simpler solution, at least superficially. If somebody manages to break both wireguard and ssh security, they probably had access to my laptop anyway so I've lost already.
If one is so worried about zero days in openssh, the more practical solution would be centered around these 2 questions:
1. "would you know if you got breached?" 2. "would you have any reaction time?"
A simple solution that answers this: https://github.com/64mb/ssh-login-alert-telegram/blob/master...
So the best solution according to the article is to use UDP’s feature of not being sent to any port in particular. In ideal hypothetical it should work, but on any public WIFI, and remote networks would not deliver UDP.
UDP was designed as no-guarantee of delivery. Why would you rely on it for this important feature?
Glad to see fwknop mentioned. Back when the idea of port knocking emerged, there was a lot of criticism about it. Then this came out, and not many seemed to notice.
This was before WireGuard and Tailscale, so the main option for remote access was IPsec or OpenVPN, which are both more complicated than most people want to deal with.
My checklist to secure a VPS, enforced by an Ansible playbook.
* Disable SSH root login
* Disable password login and use only certificates
* Enable fail2ban
This have been worked for more than ten years and never been hacked.
I never understand the need for port knocking.
Looks like a good way to accidentally lock yourself out
This is a lot of complexity compared to just not having ssh open to the world (on whatever port you choose to use).
Restrict it to the networks where authorized users will be connecting.
Change the sshd port to a random port of your choosing. Cuts down 99.9% of the attempts.
I've been using nftables for port knocking for a while now. I run an SSH tunnel server that needs to be globally accessible. But I don't want it getting hammered by bots nonstop.
So, I have this nft script which works alongside Firewalld:
$ systemctl enable --now nftables
$ cat /etc/nftables/portknock.nft
table ip portknock {}
delete table ip portknock
table ip portknock {
set knocked {
type ipv4_addr
flags timeout
timeout 6s
gc-interval 2s
}
# Before conntrack: record the knock, then drop the packet.
chain prerouting_knock {
type filter hook prerouting priority raw; policy accept;
tcp dport 12334 fib daddr type local tcp flags syn counter add @knocked { ip saddr } drop
}
# Decision chain for port 41444. Every branch is counted so that
# `nft -a list table ip portknock` shows which path traffic took.
chain gate_41444 {
# Established/related sessions pass unconditionally.
ct state established,related accept
# Host-local. Rarely matches: host-originated traffic is DNATed in
# the output hook before it reaches prerouting. Kept as a safeguard.
iifname "lo" counter accept
# Podman containers reaching the published port (hairpin).
ip saddr 10.88.0.0/16 counter accept
# Trusted subnets.
ip saddr { 10.0.0.0/24, 10.1.0.0/24 } counter accept
# Knocked within the last 6 seconds.
ip saddr @knocked counter accept
# Default deny. If THIS counter is 0 and the accept counters are
# also 0, the chain is not being reached at all -- investigate.
# Do not assume the gate is working just because nothing got in.
counter drop
}
chain prerouting_gate {
type filter hook prerouting priority mangle; policy accept;
tcp dport 41444 fib daddr type local jump gate_41444
}
}
$ ssh -p 12334 -o ConnectTimeout=1 "${sServer}" &> /dev/null
$ sleep .5
$ ssh -o 'ExitOnForwardFailure=yes' -o 'StrictHostKeyChecking=no' -o 'LogLevel=ERROR' -fp 41444 -R "${iPort}:localhost:22" -T "${sServer}" "sleep 14d"
Yeah, it doesn't have all the "cryptographic signatures" of the article; at the same time, it doesn't have some "random" 3rd-party application that faces the internet and directly controls firewall rules that way.
It's still an OpenSSH server with key-auth only. I'm not worried about someone carefully watching my traffic and finding it. I just need Internet bots not connecting to it a million times a second.
SSH available only via VPN is simpler and more extensible.
i cannot trust tech advices from a guy who puts three ad banners (left, right and bottom, taking approx quarter of my screen space) on his personal blog.
This is what spiped is for.
With all the world's knowledge freely available on the web, people still do this.
Unless it's an April Fool joke, in June.
I don't close SSH. I just use port knocking on my firewall to enable the SSH.
AI slop article based on a flawed premise. It's 2026, the process of correctly securing an SSH server has exactly two steps:
1. Disable password auth, only public key auth should be enabled
2. Block public access to SSH entirely, use a VPN instead (Tailscale & co. make this trivial)
And 2 is entirely optional for most people reading SSH guides who just want a server to host their hobby project. Let's be real, you're probably not reading the auth logs anyway so they don't need to be clean, and if someone discovered an OpenSSH public key auth bypass vulnerability, they absolutely wouldn't waste it on you. Just let those dumb scanners go at it all day, they're not getting in.
And a little tangent: fail2ban is 100% placebo and does nothing except clean up the logs a bit. I don't understand why it's still a common recommendation for beginners, it's a relic from the past when bruteforcing was still a concern because people used password auth.
I was expecting to hear tailscale TBH. You do you though.
AI writing, sigh...
> to be unreachable: no banner, no version string,
> It works, but it has a real weakness:
Why not just use ssh as the knock protocol too? To a bespoke ssh server. Ssh to 7000, type "mellon", and ssh 22 opens up. No other software required, and you clearly already have ssh.
You run ssh?
Stacking is cool, but what happened to also having a firewall in front of it?!