I don't see how Claude helped the debugging at all. It seemed like the author knew what to do and it was more telling Claude to think about that.
I've used Claude a bit and it never speaks to me like that either, "Holy Cow!" etc. It sounds more annoying than interacting with real people. Perhaps AIs are good at sensing personalities from input text and doesn't act this way with my terse prompts..
The reliance on LLMs is unfortunate. I bet this mystery could gave been solved much quicker by simply looking at the packet capture in Wireshark. The Wireshark dissectors are quite mature, SSH is covered fairly well.
I'm anti-LLM in most cases, but:
> I bet this mystery could gave been solved much quicker by simply looking at the packet capture in Wireshark.
For some people who are used to using Wireshark and who know what to look for, probably yes. For the vast majority of even technical people, probably not.
In my case, I did a packet capture of a single keystroke using tcpdump and imported it into Wireshark and I get just over 200 'Client: encrypted packet' and 'Server: encrypted packet' entries. Nothing useful there at all. If I tcpdump the entire SSH connection setup from scratch I get just as much useful information - nothing - but, oddly, fewer packets than my one keystroke triggered.
So yeah, I dislike LLMs entirely and dislike the reliance on LLMs that we see today, but in this case the author learned a lot of interesting stuff and shared it with us, whereas without LLMs he might have just shrugged and moved on.
And thats a huge downside when people howl about "Encryption everywhere! ".
Try debugging that shit. Thats right, debugging interfaces aren't safe, by some wellakshually security goon.
You want a real fun one to debug, is a SAML login to a webapp, with internal Oauth passthrough between multiple servers. Sure, I can decrypt client-server stuff with tools, but server-server is damn near impossible. The tools that work break SSL, and invalidate validation of the ssl.
Yes, Esri products suck. Bad.
Unfortunately with SSH specifically, the dissectors aren't very mature - you only get valid parsing up to the KeX completion messages (NEWKEYS), and after that, even if the encryption is set to `none` via custom patches, the rest of the message flow is not parsed.
Seems because dumping the session keys is not at all a common thing. It's just a matter of effort though - if someone put in the time to improve the SSH story for dissectors, most of the groundwork is there.
Way to gatekeep. God forbid people use tools to help them investigate instead of knowing the exact approach to take.
My thoughts exactly. The OP used AI to get a starting point to their investigation, then used their skills to improve their game, with actual (I guess according to the article itself) proof of that, as opposed to just approving changes from the LLM.
This looks like an actual productivity boost with AI.
How much are you staking on that bet?
Sigh.
I'm still waiting for a systems engineering tool that can log every layer, and handle SSL the whole pipe wide.
Im covering everything from strafe and ltrace on the machine, file reads, IO profiling, bandwidth profiling. Like, the whole thing, from beginning to end.
Theres no tool that does that.
Hell, I can't even see good network traces within a single Linux app. The closest you'll find is https://github.com/mozillazg/ptcpdump
But especially with Firefox, good luck.
> Obviously forking go’s crypto library is a little scary, and I’m gonna have to do some thinking about how to maintain my little patch in a safe way
This should really be upstreamed as an option on the ssh library. Its good to default to sending chaff in untrusted environments, but there are plenty of places where we might as well save the bandwidth
Threats exist in both trusted and untrusted environments though.
This feels like a really niche use case for SSH. Exposing this more broadly could lead to set-it-and-forget-it scenarios and ultimately make someone less secure.
Yes, but I wouldn't be surprised if the change is rejected. The crypto library is very opinionated, you're also not allowed to configure the order of TLS cipher suites, for example.
In my experience, I call them "Suck-u-rity engineers".
They're the "wellakshually" types, with "what-ifs" of increasingly inane and stupid conditions. And add in the CVE seekers looking for anything to grab on to... Yeah, thats why they're not security, but suck-u-rity.
Ive got users to defend. And sure, security is absolutely a part. But for those diminishing returns without a real exploit, I'm not doing those. Or if I'm required by some regimen, I'm doing those in the least impact to my userbase. They have enough to fight with tech-wise without me interfering in their honest work.
+1... Given how much SSH is used for computer-to-computer communication it seems like there really should be a way to disable this when it isn't necessary.
It looks like it is only applied for PTY sessions, which most computer-computer connections wouldn't be using.
https://github.com/openssh/openssh-portable/blob/d7950aca8ea...
In practice I've never felt this was an issue. But I can see how with extremely low bandwidth devices it might be, for instance LoRa over a 40 km link into some embedded device.
Very interesting, I hadn't heard of this obfuscation before so it was well worth clicking.
Another good trick for debugging ssh's exact behavior is patching in "None" cipher support for your test environment. It's about the same work as trying to set up a proxy but lets you see the raw content of the packets like it was telnet.
For terminal games where security does not matter but performance and scale does, just offering telnet in the first place can also be worth consideration.
It made the front page when it was added.
In 2023, ssh added keystroke timing obfuscation. The idea is that the speed at which you type different letters betrays some information about which letters you’re typing. So ssh sends lots of “chaff” packets along with your keystrokes to make it hard for an attacker to determine when you’re actually entering keys.
Now that's solving the problem the wrong way. If you really want that, send all typed characters at 50ms intervals, to bound the timing resolution.
Typing with an extra 50ms latency will be fairly unpleasant.
Average is 25ms. Just put sending on a clock.
Also considering ssh tunnels.
> send all typed characters at 50ms intervals
Wouldn't this just change the packet interval from 20ms to 50ms? Or did you mean a constant stream of packets at 50ms intervals, nonstop?
I think the idea behind the current implementation is that the keystrokes are batched in 20ms intervals, with the optimization that a sufficiently long silence stops the chaff stream, so the keystroke timing is obfucated with an increased error bar of 20ms multiplied by number of chaff packets.
Whatever hair-brained suck-u-rity idiot thought this was great is an idiot.
A trick Ive used to find hidden wireless cameras is to load up a wifi pineapple or similar, and log how many encrypted packets there are.
Now, move around. If you see an encrypted spike of packets, guess what? You found a hidden wifi camera!
Same thing with ssh. If you see 33 byte spam, guess what - nobody's typing. You see different or more than the norm? Someone's typing! GASP THE HORROR! TYPING!
Now where's the real exploit here, with knowing someone's typing? Sounds like CVE inflation.
You can also use TCP_CORK to reduce the number of packets without any increased latency.
Disabling TCP_NODELAY would also reduce number of packets + be portable & simpler to implement - but would incur a latency penalty.
Haven't heard of TCP_CORK, very interesting.
For people who don't feel like googling it:
1. You TCP_CORK a socket
2. You put data into it and the kernel buffers it
3. If you uncork the socket, or if the buffer hits MSS, the kernel sends the packet
Basically, the kernel waits until it has a full packet worth of data, or until you say you don't have any more data to send, and then it sends. Sort of an extreme TCP_YESDELAY.
See https://catonmat.net/tcp-cork for where I learned it all from.
Oh wow - I've never heard of TCP_CORK before. Without disabling pings I'd still pay the cost of receiving way more packets, but maybe that'd be tolerable if I didn't have to send so many pongs. This is super handy; excited to play around with it.
I am aware of TCP_NODELAY (funny enough I recently posted about TCP_NODELAY to HN[1] when I was thinking about it for the same game that I wrote about here). But I think the latency hit from disabling it just doesn't work for me.
> That 20ms is a smoking gun - it lines up perfectly with the mysterious pattern we saw earlier!
Speaking of smoking guns, anybody else reckon Claude overuses that term a lot? Seems anytime I give it some debugging question, it'll claim some random thing like a version number or whatever, is a "smoking gun"
Yes! While this post was written entirely by me, I wouldn't be surprised if I had "smoking gun" ready to go because I spent so much time debugging with Claude last night.
It's interesting how LLMs influence us, right? The opposite happened to me: I loved using em dashes, but AI ruined it for me.
I used to love using em dashes.
I still do - but I used to, too.
I still love using emdashes, and people already thought I was a robot!
Soon the Andy 3000 will finally be a reality...
Reminds me of ethimology nerd's videos. He has some content about how LLMs will influence human language.
Some day in the future we will complain about AIs with a 2015 accent because that’s the last training data that wasn’t recursive.
The "maybe" of yesterday is the "you're absolutely right!" of tomorrow.
shouldn't it be "human language influences human language"?
ChatGPT too. And "lines up perfectly" when it doesnt actually line up with anything
Same with Gemini.
You can absolutely see this pattern in Gemini in 2026.
Btw, is the injection of "absolutely" and "in $YEAR" prevalent in other LLMs as well, or is it just in Gemini's dialect?
It's just Gemini. I'm guessing they changes the system prompt for the new year or something, but it's pretty annoying.
"You're so right, that nice catch lines up perfectly!"
It's not just a coincidence, it's the emergence of spurious statistical correlations when observations happen across sessions rather than within sessions.
You can add an M-dash, and we completed the bs-bingo. :)
I chuckled out loud. It's funny cause it's true.
Or the "Eureka! That's not just a smoking gun, it's a classic case of LLMspeak."
Grok, ChatGPT, and Claude all have these tics, and even the pro versions will use their signature phrases multiple times in an answer. I have to wonder if it's deliberate, to make detecting AI easier?
A computational necromancer has likely figured out a way to power a data center by making Archimedes spin in his grave very fast.
I'm working on a little SRE agent to pre-load tickets with information to help our on-call and I'm already tired of Claude finding 'smoking guns'.
I've love to delve into that.
Without knowing how LLM's personality tuning works, I'd just hazard a guess that the excitability (tendency to use excided phrases) is turned up. "smoking gun" must be highly rated as a term of excitability. This should apply to other phrases like "outstanding!" or "good find!" "You're right!" etc.
They love clichés, and hate repeating the same words for something (repetition penalty) so they'll say something like "cause" then it's a "smoking gun" then it's something else
You might see certain phrases and mdashes ;-) rather often, because … these programs are trained on data written by people (or Microsoft's spelling correction) which overused them in the last n years? So what should these poor LLMs generate instead?
I don't think claude has even once used this in my conversations (Claude Desktop, Claude Code, Voice conversations...) Sycophancy, yes absolutely!
Maybe it has something to do with your profile/memories?
Yes, it’s kind of a corpus delicti. ;)
At this I'm just so glad that "you're absolutely right!" phase is over.
I see it from GPT5 too a lot
smoking gun, you're absolutely right, good question, em dash, "it isn't just foo, it's also bar", real honest truth, brutal truth, underscores the issue, delves into, more em dashes, <20 different hr/corporate/cringe phrases>.
It's nauseating.
You might find this a fun read: https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing
It's what they read on The Internets when training, so don't expect them to generate new phrases, other than what they learned from it?
That's the point though, it doesn't reflect human usage of the word. If delve were so commonly used by humans too, we wouldn't be discussing how it's overused by LLMs.
### The answer that fits everything (and what to do about it)
Maybe we need a real AI which creates new phrases and teaches the poor LLMs?
Looking back we already had similar problems, when we had to ask our colleagues, students, whomever "Did you get your proposed solution from the answers part or the questions part of a stackoverflow article?" :-0
cant wait for chatgpt to make me read about grandmas secret recipe and scroll through 6 ads to see the ingredients for my chicken teriyaki dinner
Come on...haven't we all had to deal with the crazy smart lead who was loaded with those same types of annoying tics?
Considering what these LLMs bring to the table, I think a little tolerance for their cringe phrases is in order.
It's a smoking gun of Claude usage.
> Speaking of smoking guns
Oh shoot! A shooting.
So the TL;DR of this post is: don't change this setting unless you know what you're doing.
Chastise it with a reminder that you're using smokeless powder.
> I am working on a high-performance game that runs over ssh.
Found your problem.
But it is an interesting world where you can casually burrow into a crypto library and disable important security features more easily than selecting the right network layer solution.
the obtuseness is the point! This is true of a lot of my work[1][2][3].
The problems you run into when doing things you shouldn't do are often really fun.
[1] https://news.ycombinator.com/item?id=42342382
This is hackernews not consumer news
You should feel free to explore / abuse all options :)
Yea UDP is technically more performant, but then you need a crypto layer + reliable message delivery layer + bespoke client. Using a plain old SSH client is cool.
However, there are existing libraries for exactly this use case - see https://github.com/ValveSoftware/GameNetworkingSockets
I guess QUIC libraries would also work.
its not really a question of 'udp performs better'. in tcp we have to live to head-of-line blocking on losses and congestion control. if you don't care about receiving every packet, but only the most recent, then udp is a good choice.
running without congestion control means that you avoid slowstart. but at a certain rate you run into poorly defined 'fairness' issues where you can easily negatively impact other flows. past that point, you can actually self-interfere and cause excessive losses for yourself.
quic uses congestion control, but uses latency estimates and variance as a signal to back off. it still imposes an ordering on a per-stream basis. so it might not be ideal either.
sctp has a mode which supports reliable and unordered, which might be something to consider
so really - if you care about latency and have a different reliability model, its worth unpacking all these considerations and using them to select your transport layer or even consider writing a minimal one yourself
>in tcp we have to live to head-of-line blocking on losses and congestion control.
Is this not a performance consideration?
Either way, using plain old SSH means a metric bajillion computers have a client for your game built in.
I wonder if this is the same reason why Microsoft's Remote SSH plugin on VS Code is so flaky even with a decent internet connection. Every couple of months I try to give it another go and give up due to the poor keyboard latency I inevitably experience. And the slow reconnects whenever I glance away from my computer monitor briefly. This is on a fiber connection with a 20ms ping to the remote machine.
You surely mean the latency in its embedded terminal and not the code editor, right? I use VSCode’s remote SSH specifically so that code editing doesn’t suck. It really does not.
The really mysterious part is how ~10,000 packets per second costs ~20% of a core. That would mean SSH is bottlenecking in its code at ~50,000 packets per second per core which would be ~500 Mbps per core (assuming full packets) which is ludicrously slow. It is trivial to do 10x that packet per second rate. Is SSH really that poorly designed?
> It is trivial to do 10x that packet per second rate.
When making this statement, are you taking into account that SSH encrypts the traffic by default?
I do not know where people get the idea that encryption is that slow. Standard AES hardware acceleration instructions do ~25 Gbps per core (on a 2023 CPU) which is ~50x that rate [1]. I have heard modern cores can do ~40-50 Gbps, but I have not been able to find any independent benchmarks of that. Even the Intel i5-2500, a CPU from 2011, averages ~10 Gbps which is ~20x that rate. Even unaccelerated encryption can do ~2-5 Gbps in pure software which is 4-10x the SSH rate.
And in this situation, the amount of encrypted payload in each packet is 36 bytes which is ~40x less than a full packet of ~1500 bytes. You would almost surely hit packet per second limits before you hit payload throughput limits at these small sizes.
Encryption is slow when compared to data throughput you can get with a properly designed transport stack, but that is because it is in comparison to 100 Gbps per core even with no hardware offload. Anything less than ~10 Gbps/1 million packets per second (ignoring other bottlenecks, so only the software transport is the limit) is not merely unoptimized, it is pessimized.
I enjoyed this write up as it touched on several topics I enjoy reading about.
Also I was unfamiliar with SSH being vulnerable in the past to keystroke timing!
https://news.ycombinator.com/item?id=37307708
2023 discussion about it here.
> Keystroke obfuscation can be disabled client-side.
please never do that (in production)
if anyone half way serious tries they _will_ be able to break you encryption end find what you typed
this isn't a hypothetical niche case obfuscation mechanism, it's a people broke SSH then a fix was found case. I don't even know why you can disable it tbh.
That doesn't sound right to me. This obfuscation isn't about a side-channel on a crypto implementation, this is about literally when your keystrokes happen. In the right circumstances, keystroke timing can reduce the search space for bruteforcing a password [1] but it's overstating to describe that as broken encryption.
[1] https://people.eecs.berkeley.edu/~daw/papers/ssh-use01.pdf
They literally explain the mechanism in the post and then explain why the security tradeoff made sense for their ssh game………
It is to prevent timing attacks but there are many ssh use cases where it is 100% computer to computer communications where there is no key based timing attack possible.
There is an argument that if:
- you are listening to an SSH session between devices
- and you know what protocol is being talked over the connection (i.e. what they are talking about)
- and the protocol is reasonably predictable
then you gain enough information about the plaintext to start extracting information about the cipher and keys.
It's a non-trivial attack by all means but it's totally feasible. Especially if there's some amount of observable state about the participants being leaked by a third party source (i.e. other services hosted by the participants involved in the same protocol).
I agree it is more nuanced than a simple 'good for computer-to-computer' and 'bad for person-to-computer'. I'm sure there are cases where both are wrong but I don't think that necessarily changes that it makes a reasonable baseline heuristic.
this only works for manually typed text, not computer to computer communication where you can't deduce much from what is being "typed" as it's not typed but produced by a program to which every letter is the same and there is no different delay in sending some letters (as people have when typing by hand)
I haven't given this more than 5 seconds of thought, but wouldn't it make sense to only enable the timing attack prevention for pseudo-terminal sessions (-t)?
The fix seems kind of crazy though, adding so much traffic overhead to every ssh session. I assume there's a reason they didn't go that route, but on a first pass seems weird they didn't just buffer password strokes to be sent in one packet, or just add some artificial timing jitter to each keystroke.
I'm just guessing but this chaff sounds like it wouldn't actually change the latency or delivery of your actual keystrokes while buffering or jitter would.
So the "real" keystrokes are 100% the same but the fake ones which are never seen except as network packets are what is randomized.
It's actually really clever.
SSH has no way of knowing when a password is being typed. It can happen any time within the session after SSH auth.
But they'd have to be on the same network as me to do that attack, right?
Yep, like ECHELON and friends are. The metadata recorded about your (all of our) traffic is probably enough to perform the timing attack.
[delayed]
I find it disturbing.
One thing you notice if you have ADSL is that some services are built as if slower connections matter and others are not. Like Google's voice and audio chat services work poorly but most of the others work well. Uploading images to Mastodon, Bluesky, Facebook, LinkedIn, Instagram and Nextdoor is reliable, but for Tumblr you have to try it twice. I don't what they are doing wrong but they are doing something wrong and not finding out what they're doing wrong because they're not testing and they're not listening to users.
Nobody consulted me about their decision not to run fiber by my house. If some committee decides to make ssh bloated they are, together with the others, conspiring to steal my livelihood and I think it would be fair for me to sue them for the $50k it would take to run that fiber myself.
It's OK if you work for Google where there is limitless dark fiber but what about people in African countries?
It's the typical corporate attitude where latency never matters: Adobe thinks it is totally normal that it takes 1-5s for a keystroke to appear when you are typing into Dreamweaver.
I agree with your general point that most companies/projects do a terrible job optimizing for slow computers/networks, but OpenSSH is from the OpenBSD people, who are well-known for supporting ancient hardware [0]. Picking a random architecture, they fully support a system with only 64MB of memory [1], and the base install includes SSH. So I suspect that OpenSSH is fairly well tested on crappy computers/networks.
There's a good chance you have other options. Regardless of how you feel about the company's head, Starlink would probably be one of them, with likely better performance than you're dealing with on ADSL.
But you cannot just sue a company because their network connected software doesn't work well on slow networks. Let alone a project like OpenSSH. It would be like me suing a game studio because my PC doesn't meet their listed minimum requirements to play the game.
Hey, it is one thing to buy a new computer, it is another thing to ask people to move.
A better analogy is a bank redlining neighborhoods. The cost to run fiber to difficult rural locations pays itself easily if you look at a 25-year time span and is an order of magnitude less than building a new housing unit on the West Coast.
> One thing you notice if you have ADSL
This is funny to me, because ADSL used to be the fast thing, as opposed to dialup modems.
The openssh team does not owe you anything.
If you want a “1990s” mode, add it yourself or pay some to do it for you.
You're not ok with a security/privacy tool using defensive techniques because of ... the lack of fiber in Africa?
My backyard but people will take Africa more seriously than anywhere in the US 2 miles from the end of cable.
You just opened a huge nostalgia portal, never thought that Dreamweaver would still be around, I used that somewhere around 2003 I believe. Good memories
Frankly I wish there was an HTML editor that delivers on what it promised. I mean, markdown is almost as rife with edge cases as YAML and somehow the link syntax still eludes me. If we could “just” template by merging at the DOM level and had decent HTML editors the world would be a different place. But yeah, Adobe probably thinks Dreamweaver isn’t worth maintaining just as they seem to think Photoshop is barely worth maintaining (they keep adding AI features that sorta work but the foundations seem to be much worse than Illustrator)
@eieio: whatever email protection you're running is triggering on the extension info. For example I see:
> And they’re sent to servers that advertise the availability of the [email protected] extension. What if we just…don’t advertise [email protected]?
Is it possible that this is on your end?
The extension is "ping@openssh.com." It shows up in the blog reliably for me across several browsers and devices.
No, it's Cloudflare munging the HTML. Cloudflare then provides JavaScript to un-munge it, but that's not reliable.
TIL! I'll see if I can change that.
>very confidently told me that my tcpdump output was normal ssh behavior:
I mean, for modern version of Openssh it's not exactly wrong. The failure was to tell you why that is the normal behavior.
If security doesn’t matter then why not use telnet or something else besides ssh instead of forking a security library?
Telnet nowadays typically isn’t available by default for security reasons, and OP wants people to be able to play the game just by typing “ssh thegamehost”.
> I am working on a high-performance game that runs over ssh.
WAT. Please no.
Why not? If it's high-performance, it's fine.
Performing with highly elevated privileges? (Joke)
ssh the protocol doesn't imply any privileges of any kind