Soatok's Informal Guide to Threat Models
soatok.blog
soatok.blog
My introduction to threat modeling was from this post: https://www.privacyguides.org/en/basics/threat-modeling/
It's a bit shorter and focused for people interested in privacy.
Articles about cybersecurity gets 100% credibility when made by furries.
Signal pisses me off in a lot of ways.
If someone joins a group chat and posts horrific content, the admins cannot clean it up. This extremely basic functionality doesn't meet the most basic bar for group moderation and safety tools. This means a troll posting a high-frequency flashing GIF to a group chat full of epileptic people is going to cause real harm. This means someone joining a chat and posting unsolicited CSAM will legally imperil everyone present and the admins are powerless to intervene at all. They seem really indifferent on fixing this.
I would love for an alternative app to materialize that provided the same level of cryptographic excellence as Signal but without the enormous ego of their marketing teams or evangelists, which actually put a microgram of care into user experience and community safety. None of the alternatives people raise meet the bar, and I find it extremely disingenuous when people insist their privacy (which is a second-order property from their cryptographic implementations) is somehow "better than Signal". So when people do this, I tend to 0day their favored apps.
https://soatok.blog/encrypted-messaging-apps/
We, collectively, as an industry, should be able to do better. That we haven't is depressing.
I have my own ideas about requirements, but they're not concrete enough to say "requirements analysis done, let's start programming"; and most people I talk to haven't thought about this enough to be helpful.
No dice on either approach.
no, but unlike a computer, the real world isn't binary, and recognising that it's flawed and full of compromises generally heightens your chances of affecting it (by your ideas or actions).
> I'm not a fan of Signal's threat model […] but Signal's main protocol seems pretty solid, especially compared to some other systems.
My main gripe with Signal is that no amount of protocol sophistication can undo the problems linked to it being a centralised service. Soatok seems unable to acknowledge that centralisation is a real (privacy, security, reliability, political, …) concern here, nor to see value in the decentralised (federated/P2P) alternative protocols implementing the same double-ratched/PFS crypto primitives.
I genuinely do not understand where this impression is coming fron. The only thing I've ever written about this topic acknowledges that centralization has risks, but a perfectly decentralized system that doesn't properly encrypt data end-to-end is bad for user privacy.
The cryptography needs to be excellent. "But decentralization" doesn't cut it.
https://soatok.blog/2025/07/09/jurisdiction-is-nearly-irrele...
Disagreeing with me is one thing, but claiming I seem "unable to acknowledge" anytbing is dishonest.
I don't want to engage in a citation battle, I just can't care enough for that. Having read those posts about Matrix, XMPP (OMEMO) and a couple others, many months/years ago, they really came across as "screw those amateurs for even trying, Signal is great, and by my very definition of it, only Signal can be".
Again, those are not your words, but something about the tone and the way you compare them made it sound that way.
Also, even if that's besides the initial point, I firmly disagree with the premise of the post you just linked. For the same reason mentioned in a sibling comment stating that the real world isn't binary even though IP addresses might be: A centralised service is political no matter what. If not their admins, their hosing provider or executive power may decide to censor you based on your country of origin, political beliefs, ideological activism or any other reason out of your control. Signal's crypto protects what's in the envelope, but does little else (neither can it) against a motivated state-actor fingerprinting you beyond the service boundaries, and guess what, we know it to be a fact for the jurisdiction Signal is operating under.
Supoose we're trying to decide between two services for a long term group chat.
Service A, on the server-side, sees all messages, in plaintext, sent to/from all participants--including other servers. It can log it indefinitely. It sees the whole social graph. Some servers have no k-anonymity (self-hosted, single user), some have thousands of users. They're all over the world, including in jurisdictions the NSA's TAO can operate.
Service B can only see IP addresses and ciphertext. There's only one real 'server", but it has millions of users and the encryption is widely reputed by experts. Its servers happen to be hosted on American cloud providers.
By firmly disagreeing with the linked post, you are saying you prefer Service A on the matter of privacy, only because of the jurisdiction.
Is that really the hill you choose?
yep, it's almost like "Open Source" "non profit" and "we are the good guys, trust me bro" doesn't matter nearly as much as "open, federated, standardised protocols".
I've been self-hosting Jabber/XMPP for a very long time, now. There's nothing I feel I'm missing out from signal and others. Not to sound like I'm preaching, I generally tell the curious to give ejabberd a spin.
> Hybrid PQ+ECDH is a hedged bet against an algorithm break before Q-Day, but is utterly fucking useless over Pure PQ once Q-Day occurs.
> there is also the likelihood that Q-Day never arrives, either because something we don't know prevents the construction of sufficiently large quantum computers (eg. quantum gravity)
That is possible, but given the recent 2029 timelines from large Internet providers, I think it's prudent to prepare for Q-Day even if it never arrives.
> or because the entire field was a scam.
The field is like... a magnet for scams, sure. But it, itself, isn't one.
And, like, the Quantum Village at DEFCON has really failed to establish credibility in my eyes.
https://soatok.blog/2022/08/18/burning-trust-at-the-quantum-...
https://soatok.blog/2023/08/20/defcon-quantum-village-2-elec...
> in that scenario abandoning ECC would have been pretty stupid.
Not really, no. See https://blog.trailofbits.com/2024/07/01/quantum-is-unimporta... for a counter-point.
> That is possible, but given the recent 2029 timelines from large Internet providers, I think it's prudent to prepare for Q-Day even if it never arrives.
the article you linked supports my position.
> the fear of the quantum doomsayers is based on a completely valid observation: the internet has put nearly all of its cryptographic eggs into the single basket of the hidden subgroup problem.
> By the time the next phase of standardization is over, we can expect to have algorithms based on at least three or four different mathematical problems. If one of the selected problems were to fall to advances in quantum or classical algorithms, there are readily-available replacements that are highly unlikely to be affected by attacks on the fallen cryptosystems.
Where did I ever make that argument? In both TFA and my previous blog post, I've made it abundantly clear that I'm pro-hybrid.
My argument is simply:
1. The claimed benefits of ECDH hybridization evaporate immediately the moment Q-Day happens. No one disputes this.
2. Harvest Now, Decrypt Later (HNDL) is the primary threat we face today during the uncertain times where we don't know if Q-Day will ever happen.
Advocating for PQ+ECC hybrids over PQ is fine. But fear-mongering about PQ in this threat model is self-defeating: Once Q-Day happens, your only source of security is PQ anyway, so if we're going to do hybrids with today's threat model in mind, PQ+PQ is the way you really want to go (and PQ+PQ+EC if you really want EC). The blog post you're commenting on says this explicitly.
I'm not anti-hybrid. I'm anti "this is an NSA ploy" bullshit. And the IETF mailing list thread I'm mentioning is stuffed with this kind of irritating conspiracy theory rhetoric. I even link to, and quote, two examples of this.
it wouldn't even occur to me that someone would take time addressing it without being one of those anti-hybrid people.
It’s like watching a cybersecurity version of Dawsons Creek or The Young and the Restless or… Jerry Springer?! Insane
I want to broadly agree but I still can't resist arguing :)
EC is really cheap on the CPU and I trust that libsodium's X25519 is implemented pretty solidly. After Q day, the $ price to break EC is still not negligible.
Whereas PQ+PQ is really expensive. I'm anti PQ+PQ hybrid just on cost. PQ+EC is practically free and still inflicts $'s on attackers after Q day (attacks do get cheaper and you discard the EC at some point, but practically I don't see EC as instantly worthless).
Your wording ("Once Q-Day happens") strongly suggests Q-Day will happen, like, it’s so certain you don’t even need to state it explicitly, you can just assume it will. And your references to the PQ timeline give the impression that you think it will likely happen soon.
It’s pretty clear from there that you think ECDH is now technically useless, and the only real justification for hybrid schemes (as opposed to pure PQ), is to reassure the people still unsure about the likes of ML-KEM. Sure you still do recommend going hybrid, but from what I can tell, you would have preferred a world where we go pure PQ right away.
And so would I to be honest (if ECC is a bust): one algorithm is simpler and faster than two.
What does it matter that my public arguments are tactical? Hybrid gets us to PQ faster, which makes progress on plugging up the HNDL risk.
> Your wording ("Once Q-Day happens") strongly suggests Q-Day will happen, like, it’s so certain you don’t even need to state it explicitly, you can just assume it will.
The literal opening section is talking about recent changes in direction from large Internet providers about quantum computing risks.
The rest of the article is predicated on "these companies' risk assessment turns out to be correct".
Separately, in https://soatok.blog/2024/09/13/e2ee-for-the-fediverse-update... I wrote more about my actual beliefs about the likelihood of Q-Day.
> It’s pretty clear from there that you think ECDH is now technically useless, and the only real justification for hybrid schemes (as opposed to pure PQ), is to reassure the people still unsure about the likes of ML-KEM. Sure you still do recommend going hybrid, but from what I can tell, you would have preferred a world where we go pure PQ right away.
You are extrapolating from the subsidiary clause of an if statement whose truth value I do not claim to know.
> And so would I to be honest (if ECC is a bust): one algorithm is simpler and faster than two.
Sure.
Do we have reason to suspect Google and Cloudflare have inside knowledge about quantum computers? To me this is more about the end of the NIST contest, and that one has no bearing on actual advances in quantum computing.
> The rest of the article is predicated on "these companies' risk assessment turns out to be correct".
Err, where did you wrote that? I can’t find it in your last two articles.
> You are extrapolating from […]
I exptrapolate mostly from this:
"I generally prefer hybrid KEMs–not out of any practical concern over ML-KEM’s security (or any other PQ KEMs, generally), but for reasons I’ll explain later in this blog post."
And this:
"Hybrid KEMs are an easier sell to people who are not cryptography experts than pure post-quantum KEMs for reasons that are mostly related to psychological safety than cryptographic safety."
https://soatok.blog/2026/04/13/hybrid-constructions-the-post...
Sorry if I’m misinterpreting, but as you can see I’m not the only one.
---
Anyway, good article on threat models.
Just now. In an HN comment.
I write in conversational English. I'm not always going to meticulously write everything like a formal argument might.
If you didn't understand that what I wrote later in a blog post was predicated on an assumption established in the intro, but would have if I wrote an explicit transitional sentence, that's useful feedback. But if you're treating an informal blog post like a court filing, you might be setting yourself up for disappointment.
Fair enough.
When I write an article (and to a lesser extent even a comment like here), I tend to agonise over every sentence. I’m guessing I’m kinda assuming others do the same. Except of course they don’t.
My dayjob involves a lot of code review and protocol cryptanalysis, so I agonize quite a bit there.
My blog would be less fun if I maintained the same level of rigor. If that makes any sense. ^^;
Yes.
Both have internal global security orgs that are constantly communicating with other large companies and governments. If they are accelerating (as are others), it is a signal.
The reliability of that signal is up to the reader to determine.
https://www.microsoft.com/en-us/security/blog/2026/06/30/mic...
Aaand, there’s no citation, no reference or link for further reading, no justification for the claim. Not the most reliable signal.
You're thinking like this is an academic release, it is not.
This is more like a caution signal given by multiple global corporations with close ties to intelligence agencies, in some cases being State actors for intelligence. NSA and CIA release products, but they try not to leak their methods and sources. Same applies.
The main thing I want to stress here is: I'm not anti-hybrid. Some people are. They tend to argue that less code / complexity is better, but you'll want to find one of them to ask directly.
1. A mathematical attack against the PQC candidates would also break ECC (I have no ability to judge this claim).
2. Implementation bugs also exist in classical implementations.
#2 seems questionable to me unless you think the same implementation bugs will exist in Curve25519 and whatever PQC algorithm you are using. If the concern is side-channel attacks then that is irrelevant to a HNDL attack. But for most communications the cost of a HNDL attack being executed several years minimum from now is far lower than the cost of an implementation bug in ML-KEM breaking their security today. Whereas Curve25519 is very well tested in its standard implementations.
Hybrids obviously help if you believe Q-Day is far into the future, or never coming.
But if you take Q-Day happening as possible in our lifetime, the HNDL threat means data being encrypted today depends entirely on PQ security in the long run (since breaking EC with a Quantum Computer has an attack cost of like 2^30 or so instead of 2^120 or so).
A lot of big players have accelerated their Q-day timelines fairly recently by years.