I'm confused, is the argument that it doesn't work because Google is fueled by surveillance capitalism? If so what about Apple which is only partly so? What about Firefox and in particular its de-branded ones without Google search as default?
I think what makes the Web special is precisely that there are different browsers beyond Chromium. If the Web was Chrome I would tend to agree but even though popular I do not think it is fair to conflate it to be the Web.
I could not find anything about google or other browser vendors in the article.
My take is that you should trust provider (developer, hoster) of said encryption app to send you actual implementation, not something that looks like the real deal, but does not encrypt anything. From a regular user's point of view: you can not inspect what you run (due to technical reasons, that on the web anything can be downloaded and executed at any moment, swapping implementation on the fly. And due to skills needed to actually read and understand executed scripts), so you can only believe and trust. At which point usual TLS is surely enough.
Like I said I'm confused, genuinely trying to figure the article out.
"A cryptosystem is incoherent if its implementation is distributed by the same entity which it purports to secure against."
What is the cryptosystem then on the Web? Who is the entity? It's not the server or the Website so I don't see what's left except the browser and browser vendor.
There's also a long list of government (or subpeonable) entities on your certificate trust list.
Without which TLS is not gonna work.
The article is arguing that in practice you could just send your "encrypted" communications to the browser vendor, or one of the governments on the certificate root list, or someone else in the distribution chain, and have them be the middle man. The security properties of your communications would be the same. Hence "snake oil".
Things like stapling don't change this much, or reduce to TOFU.
There are a lot of other implementations of this idea that don't necessarily rely on trust-on-first-use. The securedrop team explicitly includes malicious JS served by the primary-domain in the threat-model and made WEBCAT[0] as an outcome of that research. Their article on webcrypto is much better than this one.
The solution obviously is to go out-of-band:
> When a user visits a website that has enrolled in WEBCAT, before the site can load the content is checked against a signed manifest to ensure that it has not been tampered with (more on enrollment later). If everything checks out, the page loads normally. If, however, any content does not match what’s expected, the page load is aborted and a warning is displayed, protecting the user from potentially malicious content before it can execute.
The entire argument is based on the definition of an “Incoherent cryptosystem”, which is too restrictive to be useful for cases that you want eg. Tor is also developed and distributed by Tor people and it is supposed to protect you against everyone, including the Tor people.
You could also audit the JavaScript / Wasm that's running in your browser. In fact, a security-focused e2e application might want to purposely keep all client-side code un-minified and highly readable for this very purpose, but decompilers and LLMs could provide reasonable auditability regardless
I think the article raises interesting questions about trust, but I am also doubtful if the definition of the “incoherent cryptosystem” is useful:
The article argues that Signal is an incoherent cryptosystem, because they ship the E2E-encrypting Signal client (and could, hence, backdoor it) that should protect me, the user, against their own infrastructure snooping on me.
As I understand the definition, we would not have an incoherent cryptosystem if I used a third-party client on Signal's infrastructure. Said Non-Signal client would implement E2E encryption, and use the Signal infrastructure, so the entity running the infrastructure is different from the entity providing the client. But is this any better?
Couldn't “Non-Signal Corp.” be coerced by the government (or decide to build a backdoor for their own gain) just as easily as “Signal”?
So I don't think it matters if the entity distributing the client is the same as the one running the infrastructure. It matters if I trust the client. How to implement this (audits, OSS, pinning of version) is still an open question to me.
This is precisely why I have autoupdates disabled for my Signal apps. They're on by default, which basically gives Signal-the-org remote code execution on my machine (same as Chrome's built in transparent autoupdate gives Google RCE on your machine).
I guess you could make a point for using messengers based on open protocols (like Matrix) that have plenty of different client implementations. It doesn't protect you from targeted attacks (it might if you can somehow hide from the outside world which client you're using, or if you write and maintain the client yourself) - but it makes it less likely to be affected if your favourite agency managed to backdoor some random client implementation.
Signal seems to be quietly warning people away from the service by refusing to update their privacy policy the very first line of which is a clear lie.
Before reading this article, I used to believe that IT companies deeply respected users’ human rights, spending millions of dollars to build end‑to‑end encryption. But thanks to this very article, I learned that they were actually saving tens of millions in administrative litigation costs – costs they would otherwise have had to pay every month to respond to wiretap warrants.
Some might call this a “cryptographic innovation.” I call it “the technical outsourcing of legal disclaimers.” Unfortunately, I don’t seem to have a Harvard Law School legal team on my side.
End-to-end encryption is about protecting data at rest on the vendor's servers. TLS only secures data in transit.
The article's argument is a bit like saying TLS protects plain-text passwords in transit, so there is no need to store them in hashed form in the database.
Sure, the article makes good arguments about the trust that is still implicit in E2EE, but it goes too far in its dismissal of it.
>End-to-end encryption is about protecting data at rest on the vendor's servers
No its literally in the name. It's about encryption between the ends of data communication. In most cases between two clients with the middle man server exchanging and hosting the encrypted data between them. Its about protecting live data which by design results in at rest data protection. FDE for example only protects at rest data but does not prevent live data extraction.
As a developer, when Facebook says that WhatsApp is end-to-end encrypted, you know that it means "the default client provides encryption, by default" but you know very well that they could selectively turn it off for anyone, at any time and it may be impossible to know that they did this this unless the target was tech savvy and actively monitoring their network packets. Especially on the web, they could just serve a different JavaScript library with a backdoor for a specific IP address only and the target would have no idea.
If web-based encryption is snake oil, then science-based medicine is also snake oil, because you trust your doctor not to secretly give you sugar pills instead of the real thing. In fact, this argument applies even more strongly to medication, because I can't really determine what a pill does, but I can determine what an app or website does and what it sends to the server.
This entire article is... Nonsense? It categorically dismisses e2ee, without any supporting evidence whatsoever, other than the notion that a provider might push a update that doesn't encrypt messages anymore.
It's on the level of "you can't trust your OS unless you wrote it yourself"
-righteous sounding but utterly stupid in practice
> A cryptosystem is incoherent if its implementation is distributed by the same entity which it purports to secure against.
This is both true, and also useless: pretty much any E2E system is falling under this definition.
By definition you can't protect yourself from the entity that provides you the software you use, because you have now way to guarantee that they aren't going to backdoor you.
That doesn't mean it's snake oil though, as the entity you want protection against is generally not the software provider but a third party. Using e2e from a US-based entity means you are prone to spying from the US government, but at least you know you're reasonably secure against the IRGC, the Chinese intelligence service, the FSB, and so on.
It also means you are safe from data leaks, which are by far the most common threat today.
No system can be secure unconditionally, it's always secure under a particular threat model.
One thing you can do is have your adversary put their money where their mouth is and use the very same products, sourced independently, that they use to protect their own sensitive information.
There are limits to this of course. You can’t buy a TACLANE[1], but you can buy many of the other products[2] USG uses to protect its own classified information.
The obvious counterexample is NOBUS[0] vulnerabilities, and intentional backdoors like the Clipper Chip[1] or Dual_EC_DRBG[2]: if you genuinely believe you are the only one who could possibly exploit it, there's no reason to avoid using it.
A more modern example is probably the NSA aggressively pushing[3] for replacing classical encryption with post-quantum encryption, rather than taking the more conservative and probably-more-secure approach of layering the two - while at the same time mandating the use of two layers of those same algorithms for their own use[4]!
> The obvious counterexample is NOBUS[0] vulnerabilities, and intentional backdoors like the Clipper Chip[1] or Dual_EC_DRBG[2]: if you genuinely believe you are the only one who could possibly exploit it, there's no reason to avoid using it.
The problem with these examples is that they weren't used in national security systems, which are the systems for which NSA has a legislated defensive responsibility.
Clipper was designed for use by the public; it was not intended to ever be used to protect classified (or even sensitive unclassified) information at all.
Likewise with Dual_EC_DRBG. The CSfC component requirements drew from the Common Criteria Protection Profiles, where Dual_EC_DRBG was never an option.
The NSA isn't aggressively pushing for PQC; the industry is. Note that the PQC standard we have was the product of a competition won by European academic cryptographers.
That is not what the comment to which you replied said.
There is a consensus about pushing PQC and about the PQC standard.
The dispute is about something else, about whether the current Diffie-Hellman based key establishment algorithms should be immediately and completely replaced by PQC (which is what NSA pushes), or to be more prudent and use for some time both methods (so that the key exchange cannot be broken when any of the 2 algorithms is broken), to afford more time for gaining confidence in the standardized PQC algorithms, until eventually one might decide that it is reasonable to omit the current algorithms.
What NSA wants reduces somewhat the costs, but not by much, because PQC is much more expensive, so most of the cost is determined by it and not by the classic algorithms, even when they are used together, and also because keeping the current algorithms does not require any development work, as there are mature implementations in SW or in HW for all applications. The opponents argue that this small cost reduction is not worthwhile, because it eliminates the serious risks that flaws may be discovered later in the current PQC standards.
Moreover what NSA wants would complicate the protocols, because many would not accept the risks of the NSA variant, so it would remain optional to omit the classic algorithms, increasing the number of choices in the protocol, which is always undesirable in security protocols.
No, I'm very clear on what the question was asking, and it is not in fact "NSA" driving this process.
This is 100% Daniel Bernstein drama. Bernstein is upset that the NIST contest selected MLKEM (Kyber). He's been running a yearslong crusade to impeach the standard, up to and including opposition to lattice cryptography writ large, despite himself signing on to a lattice entrant (SNTRUP) to the PQC competition.
Over the last several months, the focal point of this crusade has been the IETF TLS Working Group, where Bernstein has been canvassing opposition to an RFC that will establish code points and documentation for pure, non-hybrid MLKEM TLS. Few systems in the near future are likely to use pure MLKEM TLS, but there are conceivable systems that we can foresee needing the option: for instance, embedded systems like smart meters that do ECC today, and could do MLKEM if it was urgently needed because of a CRQC, but couldn't practically do both.
The arguments being made on TLSWG are in spectacularly bad faith. People have varying levels of confidence in pure MLKEM (the closer your job title is to "cryptography engineer", the more likely it is you think hybrids are silly). But nobody in the entire world is being forced to use pure MLKEM; the entire controversy is about documenting what TLS would look like if you needed pure MLKEM.
The "NSA" stuff is just more innuendo.
Apart from bytes on the wire, MLKEM is not in fact "much more expensive" than 25519. Also, you have the maturity argument wrong; it's not a concern that risks/flaws will be discovered "in the standards".
Your argument would be far more charitable if NIST had not already been caught pushing a broken standard at the behest of the NSA before. DJB might be combative and somewhat caustic, but the one thing he's never been, given enough time in the retrospect to show it, is wrong.
The logic you're using here would be just as valid in a campaign against SHA2. More valid, in fact, because unlike MLKEM, which was designed by European academic cryptographers, the NSA actually designed SHA2.
> As part of its efforts to foil Web encryption, the National Security Agency inserted a backdoor into a 2006 security standard adopted by the National Institute of Standards and Technology, the federal agency charged with recommending cybersecurity standards. Credit Patrick Semansky/Associated Press
There is no source in the article to support that claim, it links to another article where it merely says that people suspected it of being backdoored, but those people had no proof either. There are objective reasons why it's not a good idea to use compared to better algorithms, but I wish the reasoning was not solely based off of a conspiracy theory or hate for the NSA.
> This is both true, and also useless: pretty much any E2E system is falling under this definition.
This is not true. I can build Signal from source from GitHub, and use Signal-the-service with the client (which did not come from Signal, but GitHub/my compiler).
Many cryptosystems are like this. In any case, if you are getting something from the App Store, you can get it once and disable autoupdates, which prevents the service provider (presuming they are the same as the people who published the app) from backdooring you at some point in the future. Alternately, even with updates, unless Apple is colluding with them to serve only you* a specific backdoored app, you can at least be reasonably confident that it's not specifically backdooring only you* in an undetectable fashion.
> This is not true. I can build Signal from source from GitHub
Sure, but can you find an NSA-designed backdoor in the source code?
> you can get it once and disable autoupdates
Try doing that with Signal, and you'll be unable to connect to the main network in just a few days because you get out of sync. Also, what do you do if there's a high severity CVE on the program? You still don't update or you re-audit all the new code?
What you describe may be possible for an intelligence agency, but completely out of reach for an individual.
> unless Apple is colluding with them
Given the most likely adversary is the US intelligence, it's absolutely not far fetched to assume that in your threat model.
> you can at least be reasonably confident that it's not specifically backdooring only you
>> This is not true. I can build Signal from source from GitHub
> Sure, but can you find an NSA-designed backdoor in the source code?
You're moving the goalpost. They were responding to the claim suggesting it's impossible to get non-Signal provided signal.
>> you can get it once and disable autoupdates
> Try doing that with Signal, and you'll be unable to connect to the main network in just a few days because you get out of sync.
That's demonstrably false. On one of my idle/backup phones I'm using Signal 8.8.2, released in April 2026, almost 3 full months ago. It can not only connect to the network but everything works, with every contact.
You might think of the official Signal client expiration, but that's client side (meaning that you can compile and use the version that doesn't have it) and..... 90 days, not "a few".
I don't have a concrete number for the server side of enforcement though (minimumVersions seems to be populated at start time, with the defaults not committed to the repo). It's not entirely unreasonable to assume that the lowest official supported version is the one that introduced the concept of usernames, and the only meaningful capability test is SPQR.
> Also, what do you do if there's a high severity CVE on the program? You still don't update or you re-audit all the new code?
I think disabling auto update was shown as a possible strategy against a silent, targeted auto update. Not a way to remain protected against the general CVEs.
> You're moving the goalpost. They were responding to the claim suggesting it's impossible to get non-Signal provided signal.
That was never my claim. The claim is that you cannot protect youself from Signal being malicious if Signal is the maker of the software. Compiling the software yourself doesn't help against the kind of adversary in the threat model.
> That's demonstrably false. On one of my idle/backup phones I'm using Signal 8.8.2, released in April 2026, almost 3 full months ago. It can not only connect to the network but everything works, with every contact.
Lucky you, you only need to fully audit the codebase every 3 months.
I'm using the Signal apk directly so I'm painfully aware of the frequency of the breakages.
> I think disabling auto update was shown as a possible strategy against a silent, targeted auto update. Not a way to remain protected against the general CVEs.
I don't think you understand my point. I'm not talking about the CVE being exploited about you. The CVE will just push you to download the compromised update, breaking your “security through lack of update” policy.
Defense against Signal may not be total, but if you build from source (or use a reproducible build) you can check to be pretty sure you have the same software as everyone else, in which case a backdoor would have to be in a public version of Signal, which means any backdoor has a chance of getting discovered.
Backdoors can be well-hidden, but this is the kind of software that people look for vulnerabilities in [1]. This gives a lot less flexibility to any potential backdoor.
Additionally, someone could decide to write a new Signal client from scratch, designed to be compatible with the original server. Such a client would probably be less secure overall, at least at first, but it couldn't have a backdoor inserted by Signal developers, since it wouldn't contain any Signal code. Since the original client also supports end-to-end encryption, a new client can be compatible with old clients.
Now the argument is shifting. It started out as "a cryptosystem can't be secured from the entity in control of its supply", and now it's "a cryptosystem can't be secured even given its source code because the NSA".
If you get an open source app from the App Store, is there any assurance it actually reflects the code in the repo? I’d think the signing step happening in isolation opens the door to tomfoolery.
> the entity you want protection against is generally not the software provider but a third party.
This. The author is dismissing the whole web-based cryptography, or any end-to-end cryptography for that matter, on the basis of a one-dimension analysis.
But claiming that your system is end to end encrypted means that you are claiming protection from you and your system. This is mainly a truth in advertising issue.
> means that you are claiming protection from you and your system.
Not necessarily. I push for e2ee everywhere I can for a completely different reason: when (not if, when) we get breached, we cannot leak sensitive data we don't have.
Only if you manage to deploy malicious code and have it stay there undetected. But in practice this is much harder than having access to a DB or a backup for a few minutes (that's all it takes to leak data).
(Just as an example, at my current employer, out of 8 people in the company, there are 8 people who have credentials that permit at least some db access in one way or the other (anyone that has at least one of this job: customer support, account managers, data scientists, devs). There are only 3 people who can push changes to production, and you need 2 out of 3 to do so, and all of three get paged when a release happen. Of course it's not infallible, but the probability of the app being corrupted is orders of magnitude lower than the DB being breached, by the mere scale of the number of people who have access to these things. And again, you only need transient access to the data to leak it, when you need persistent access to do damage by corrupting the code).
> By definition you can't protect yourself from the entity that provides you the software you use, because you have now way to guarantee that they aren't going to backdoor you.
That's not completely true. If I can control when (and if!) the software updates and if there is some kind of vetting process to verify that the version I'm currently running does not contain a backdoor, I can treat it like a third party with respect to the server.
I agree with you though that most current software that are made to auto-update at any time without any oversight do not fall under this umbrella. Web apps definitely don't fall under it.
> Using e2e from a US-based entity means you are prone to spying from the US government, but at least you know you're reasonably secure against the IRGC, the Chinese intelligence service, the FSB, and so on.
You don't need E2E for that, using https/TLS for transport and servers hosted in the US would be enough.
But that's OP's point. If the server is pwned, the hackers can simply change the front-end of the app and have it send the confidential data to wherever after it was decrypted on the client.
Anyone using reverse proxies, CDNs or anti-DDoS services already voluntarily give full MITM privileges of their unencrypted data to companies like CloudFlare, Amazon, Akamai, Fastly, etc., which is most of the top sites on the internet and a large percentage of the overall internet traffic.
>...pretty much any E2E system is falling under this definition.
The definition is quite clear. It does not apply when the implementation is not distributed by the same entity that creates it for example. There are other related issues but the message here is that web based cryptography has a particular weakness when it comes to things like end to end encrypted messaging.
> By definition you can't protect yourself from the entity that provides you the software you use, because you have now way to guarantee that they aren't going to backdoor you.
But with repro builds and system transparency, hiding backdoors is impractical.
Not really, time and time again people have abused their position/influence to build backdoors almost into everything for good or bad reasons. The whole idea of third party audits was to ensure that there are checks and balances. Then again the auditors are lowest paid to get stuff done and they take word of the company for most part.
On other hand its quite natural, security is not really getting you direct revenue so business is least motivated in investing it or say continuously investing in it. The ones that do are doing partial lip service for most part.
> Using e2e from a US-based entity means you are prone to spying from the US government, but at least you know you're reasonably secure against the IRGC, the Chinese intelligence service, the FSB, and so on.
Framing this in terms of governmental espionage is nonsensical. Using e2e from a US-based entity makes you completely sure that the US government is spying on you, because they assert direct control over the software you are running. There is no venue to seek justice against an unlawful contract if the government is in on it.
You should instead take a step back into reality and consider data misuse by normal non-government actors. Facebook claiming e2e encryption is a contractual matter, that you can litigate. That's where the actual protection is. That's also why real business demands not "unbreakable encryption" but reliable marker for access and tampering. It is much more useful to have a record of who accessed the data, than a claim that it's impossible.
> By definition you can't protect yourself from the entity that provides you the software you use, because you have now way to guarantee that they aren't going to backdoor you.
Well that’s just wrong, just make it open source and do reproducible builds
Two problems I see with the authors argument. Maybe someone more knowledgeable can chip in to correct me if I'm wrong:
1. Aren't E2EE systems designed to prevent decryption of content already created in the past sitting on the vendor's servers? Yes, the vendor could go rogue, but, assuming they currently have implemented E2EE right, it means any change to the client can only compromise content created in the future from that point onward, no? So why is the article implying Apple could have provided a back-doored iOS to bypass the encryption for existing content?
2. I also don't find the argument that E2EE is only a legal trick fully convincing. There are several other incentives for a vendor to implement it apart from avoiding legal issues: preventing insider abuse, reducing liability, improving customer trust, and resisting mass surveillance
These are real engineering motivations. The threat model is not: "Protect you if <vendor> becomes actively malicious tomorrow." Its more like "Protect messages stored on <Vendor>'s servers from attackers, employees, hackers, routine legal requests, and passive surveillance."
Alright, I'm ready. These are engineering motivations, as you said. So, which one of these isn't a cost center? Because an insurance policy would handle the first two, but probably cheaper. Customers have repeatedly proven they will buy the product lacking the trust. Resisting mass surveillance? They are the mass surveillance. Which is now a legal compliance based cost center.
The author is basically saying if you participate in any part of the encryption process, you're deceiving users in saying things are e2e encrypted.
Isn't this conflating encryption with trust? Of course whoever claims to encrypt your data needs to be trustworthy, and whether they actually are is another matter, but If my app allows you to generate a client side key, export it and use it to encrypt data client side and we only get the encrypted data, that is verifiably valid encryption.
I could be malicious and also send a copy of your actual plaintext to the server as well, but that is trivial to check (unless I'm being targeted and I am the only user that gets the malicious code, still, I can check). It's a risky proposition for an organization with vested interest in being seen as pro privacy.
But I get it, different conversation if the government coerces you, and the outcome depends on your bank account and ability to handle pressure.
If you’re not running PGP commands yourself to encrypt and read your messages, it’s not secure.
Also no OS integrated system that does this for you automatically / conveniently has ever existed that was widely adopted because that application would have the ability to read all of your private communication, and impossible to install on an uncracked phone.
Still it would take literally minutes to vibe code an app that sits in front of a WhatsApp client and automatically handles these things. Maybe the future is just to write it yourself (not the security) so you can trust it and it’s convenient.
Isn't non-web-based cryptography affected (as per this take) in the same way but with extra steps?
A sophisticated actor might as well also control the application that ends up on my device.
It does not have to be the same delivery mechanism as long as I did not write it myself.
So all cryptography is snake oil?
___
I mean I kinda sorta get the point and there would be some merit to discuss there, but the weird framing makes that very hard to do.
Of course it's easier to break web e2ee if you are for example cloudflare compared with someone also having to compromise the Debian repos.
How about GPG distributed with a Linux distribution like Debian as a counterexample? It would be fairly difficult to backdoor GPG in that case without getting caught. Everything happens in the open both at the GPG level and the Linux distribution level. The binaries are signed by the distribution and are distributed by a bunch of mirrors. An evil Debian maintainer would have to make a change that was well enough disguised as something else to evade scrutiny.
> An evil Debian maintainer would have to make a change that was well enough disguised as something else to evade scrutiny.
The xz utils hack got slurped up into sid before it was discovered by a researcher's performance regression in ssh. IIRC the hacked test file didn't even need to be added to the upstream source tree because Debian was blithely downloading release tarballs from Github. No evil Debian maintainer needed.
It's funny that when speculating about Debian's security you forget an actual state-level attack that got code into sid, but when speculating about Signal's insecurity in another thread you're quite happy to imagine potential state-level attacks.
> Isn't non-web-based cryptography affected (as per this take) in the same way but with extra steps?
Yes, but it's a whole lot of extra steps spread across multiple independent parties, each of them adds large delays to the actions and increasing the chance that it is discovered long before it ends up on the users machine.
When you hack GPG it will take years before it trickles down into every Linux distribution, especially LTS releases. And ideally, you want an encryption protocol, not one app, thus you have some people running GPG, some running Sequoia PGP and some running OpenPGP.js. If somebody fiddles with the encryption, different clients won't be able to decode the messages anymore and it will be clear pretty quickly that something is wrong.
Meanwhile on the Web or smartphones, you remove or backdoor the encryption, everybody gets auto updated to the latest version and nobody will know that something went wrong.
Many comments suggest that the definition and criteria proposed are infeasible and useless. They are not wrong.
But still it points out something the layperson misses (or even many tech-savvy people): An E2EE service from A does not provably protect your data against A.
this is one of the problems content addressed stores like nix and ipfs can prevent. every version of the code is immutable and impossible to delete. if the devs update the "latest pointer" to a backdoored release users can just stay on the old version or move to a fork. and in the happy case (honest developer) you get all the benefits of auto update.
for this to work in practice it needs to be paired with reproducible builds, open source and either p2p or server choice (use signal.mydomain.net instead of signal.org). but these are all things that already exist and none of them is really hard to set up. the harder problem is distributing community block lists of bad package versions but that can be done with atproto or simple ublock style filter files.
i think the real bottleneck for adoption is that the only browser with built in ipfs support is brave, the one thats full of crypto ads and affiliate link fraud. i dont know if firefox would ever take it up or we need to build a brand new browser. or find a way to do it one layer down with a system service.
Reminder: iMessage claims to be e2ee, but the on-by-default iCloud Backup on iOS backs up material that is sufficient to defeat this (either the endpoint keys, in the case of "Messages in iCloud" disabled, or the messages themselves, in the case of "Messages in iCloud" enabled).
This means that, in practice, iMessage is not e2ee.
Before you say "But what about Advanced Data Protection that enables e2ee for iCloud Backup?" - virtually nobody has this on, Apple prohibits you from turning it on in the UK, and even if you enable it - the people you iMessage with don't, so your conversations are in their backups. This means that if either endpoint of the iMessage conversation is in the UK, and both parties have iCloud Backup enabled (the default), then your iMessages are not e2ee as a non-endpoint has an escrowed copy of the plaintext or keys.
I feel the legal part is on point. It's also increasingly used by governments to have their cake and eat it too: "We'll completely lock down your devices and run all kinds of analysis on your data, but don't worry, it's all done on-device and all communication is encrypted, so our promise to protect your privacy is kept!"
Kind of a weird concern to be reading on HN, where the sentiment is overwhelmingly that these kinds of systems should be exempt from warrants and discovery.
If, say, Signal was completely controlled by the CIA[1] and was thus evil, then having incoherent cryptography as described in the article would be a feature, not a bug. Being able to reject law enforcement requests would produce a false sense of security for the people the CIA was interested in surveilling. Responding effectively to law enforcement requests would reduce the value to the CIA of the ability to secretly backdoor Signal.
This effect was seen in the Apple vs FBI incident described in the article. The public perception of Apple as a brave defender of their user's privacy was greatly increased due to that dispute. For all we know, the FBI was in on the conspiracy. In return they might receive the fruits of such surveillance with the only limitation that they would have to disguise the source with parallel construction.
I don’t agree with this. While, it may be true that E2E carries risks of government surveillance.
E2E makes it less likely that your information will get hacked and reduces the risk that employees will access your information.
The reality is that these security claims are generally subject to internal audit and would need company wide collusion and the risk of a whistleblower or disgruntled former employee if they were violated provides some level of protection that a large tech company offering of e2e doesn’t mean some level of benefit from the user compared to perfect encryption security.
Something I don't understand about this argument, which has been made before, is you can tie JavaScript to a specific hash with SRI. So you release the cryptographic code in public where it can be audited and then what runs in the browser verifies that was what loaded.
The host could inject malicious JavaScript from the host or change libraries but I feel like this is an avoidable problem because it can be audited much more easily than expecting users to audit JavaScript every time. People could even build known, trusted, web frontends. So I think there are mitigations if not ways to assure the browser is running trusted code.
in fact, this could be generalized more and doesn't neccessarily have to be about hiding messages. We've all heard the discussions about using VPNs "for privacy" (i. e. for hiding your IP metadata), when it's really just shifting trust away from your ISP and towards the VPN supplier.
It always comes down to who the alternative party to trust would be. In the fictional dialogue, the alternative appears to be to not send the message. Which may or may not be the better option than to give Eve eavesdropping capabilities.
104 comments
[ 0.21 ms ] story [ 55.1 ms ] threadI think what makes the Web special is precisely that there are different browsers beyond Chromium. If the Web was Chrome I would tend to agree but even though popular I do not think it is fair to conflate it to be the Web.
My take is that you should trust provider (developer, hoster) of said encryption app to send you actual implementation, not something that looks like the real deal, but does not encrypt anything. From a regular user's point of view: you can not inspect what you run (due to technical reasons, that on the web anything can be downloaded and executed at any moment, swapping implementation on the fly. And due to skills needed to actually read and understand executed scripts), so you can only believe and trust. At which point usual TLS is surely enough.
"A cryptosystem is incoherent if its implementation is distributed by the same entity which it purports to secure against."
What is the cryptosystem then on the Web? Who is the entity? It's not the server or the Website so I don't see what's left except the browser and browser vendor.
Without which TLS is not gonna work.
The article is arguing that in practice you could just send your "encrypted" communications to the browser vendor, or one of the governments on the certificate root list, or someone else in the distribution chain, and have them be the middle man. The security properties of your communications would be the same. Hence "snake oil".
Things like stapling don't change this much, or reduce to TOFU.
Your argument is a bit like saying TLS protects plain-text passwords in transit, so there is no need to store them in hashed form in the database.
https://github.com/w3c/ServiceWorker/issues/1680
The solution obviously is to go out-of-band:
> When a user visits a website that has enrolled in WEBCAT, before the site can load the content is checked against a signed manifest to ensure that it has not been tampered with (more on enrollment later). If everything checks out, the page loads normally. If, however, any content does not match what’s expected, the page load is aborted and a warning is displayed, protecting the user from potentially malicious content before it can execute.
[0]: https://securedrop.org/news/introducing-webcat-web-based-cod...
[1]: https://securedrop.org/news/browser-based-cryptography/
still this wouldn't guarantee that all the other nodes are not compromised
The article argues that Signal is an incoherent cryptosystem, because they ship the E2E-encrypting Signal client (and could, hence, backdoor it) that should protect me, the user, against their own infrastructure snooping on me.
As I understand the definition, we would not have an incoherent cryptosystem if I used a third-party client on Signal's infrastructure. Said Non-Signal client would implement E2E encryption, and use the Signal infrastructure, so the entity running the infrastructure is different from the entity providing the client. But is this any better?
Couldn't “Non-Signal Corp.” be coerced by the government (or decide to build a backdoor for their own gain) just as easily as “Signal”?
So I don't think it matters if the entity distributing the client is the same as the one running the infrastructure. It matters if I trust the client. How to implement this (audits, OSS, pinning of version) is still an open question to me.
[1] https://molly.im/
Signal seems to be quietly warning people away from the service by refusing to update their privacy policy the very first line of which is a clear lie.
Some might call this a “cryptographic innovation.” I call it “the technical outsourcing of legal disclaimers.” Unfortunately, I don’t seem to have a Harvard Law School legal team on my side.
The article's argument is a bit like saying TLS protects plain-text passwords in transit, so there is no need to store them in hashed form in the database.
Sure, the article makes good arguments about the trust that is still implicit in E2EE, but it goes too far in its dismissal of it.
No its literally in the name. It's about encryption between the ends of data communication. In most cases between two clients with the middle man server exchanging and hosting the encrypted data between them. Its about protecting live data which by design results in at rest data protection. FDE for example only protects at rest data but does not prevent live data extraction.
If web-based encryption is snake oil, then science-based medicine is also snake oil, because you trust your doctor not to secretly give you sugar pills instead of the real thing. In fact, this argument applies even more strongly to medication, because I can't really determine what a pill does, but I can determine what an app or website does and what it sends to the server.
It's on the level of "you can't trust your OS unless you wrote it yourself" -righteous sounding but utterly stupid in practice
This is both true, and also useless: pretty much any E2E system is falling under this definition.
By definition you can't protect yourself from the entity that provides you the software you use, because you have now way to guarantee that they aren't going to backdoor you.
That doesn't mean it's snake oil though, as the entity you want protection against is generally not the software provider but a third party. Using e2e from a US-based entity means you are prone to spying from the US government, but at least you know you're reasonably secure against the IRGC, the Chinese intelligence service, the FSB, and so on.
It also means you are safe from data leaks, which are by far the most common threat today.
No system can be secure unconditionally, it's always secure under a particular threat model.
There are limits to this of course. You can’t buy a TACLANE[1], but you can buy many of the other products[2] USG uses to protect its own classified information.
[1] https://gdmissionsystems.com/encryption/taclane-network-encr...
[2] https://www.nsa.gov/resources/Commercial-Solutions-for-Class...
A more modern example is probably the NSA aggressively pushing[3] for replacing classical encryption with post-quantum encryption, rather than taking the more conservative and probably-more-secure approach of layering the two - while at the same time mandating the use of two layers of those same algorithms for their own use[4]!
[0]: https://en.wikipedia.org/wiki/NOBUS
[1]: https://en.wikipedia.org/wiki/Clipper_chip
[2]: https://en.wikipedia.org/wiki/Dual_EC_DRBG
[3]: https://blog.cr.yp.to/20251004-weakened.html
[4]: https://defense-solutions.curtisswright.com/capabilities/tec...
The problem with these examples is that they weren't used in national security systems, which are the systems for which NSA has a legislated defensive responsibility.
Clipper was designed for use by the public; it was not intended to ever be used to protect classified (or even sensitive unclassified) information at all.
Likewise with Dual_EC_DRBG. The CSfC component requirements drew from the Common Criteria Protection Profiles, where Dual_EC_DRBG was never an option.
There is a consensus about pushing PQC and about the PQC standard.
The dispute is about something else, about whether the current Diffie-Hellman based key establishment algorithms should be immediately and completely replaced by PQC (which is what NSA pushes), or to be more prudent and use for some time both methods (so that the key exchange cannot be broken when any of the 2 algorithms is broken), to afford more time for gaining confidence in the standardized PQC algorithms, until eventually one might decide that it is reasonable to omit the current algorithms.
What NSA wants reduces somewhat the costs, but not by much, because PQC is much more expensive, so most of the cost is determined by it and not by the classic algorithms, even when they are used together, and also because keeping the current algorithms does not require any development work, as there are mature implementations in SW or in HW for all applications. The opponents argue that this small cost reduction is not worthwhile, because it eliminates the serious risks that flaws may be discovered later in the current PQC standards.
Moreover what NSA wants would complicate the protocols, because many would not accept the risks of the NSA variant, so it would remain optional to omit the classic algorithms, increasing the number of choices in the protocol, which is always undesirable in security protocols.
This is 100% Daniel Bernstein drama. Bernstein is upset that the NIST contest selected MLKEM (Kyber). He's been running a yearslong crusade to impeach the standard, up to and including opposition to lattice cryptography writ large, despite himself signing on to a lattice entrant (SNTRUP) to the PQC competition.
Over the last several months, the focal point of this crusade has been the IETF TLS Working Group, where Bernstein has been canvassing opposition to an RFC that will establish code points and documentation for pure, non-hybrid MLKEM TLS. Few systems in the near future are likely to use pure MLKEM TLS, but there are conceivable systems that we can foresee needing the option: for instance, embedded systems like smart meters that do ECC today, and could do MLKEM if it was urgently needed because of a CRQC, but couldn't practically do both.
The arguments being made on TLSWG are in spectacularly bad faith. People have varying levels of confidence in pure MLKEM (the closer your job title is to "cryptography engineer", the more likely it is you think hybrids are silly). But nobody in the entire world is being forced to use pure MLKEM; the entire controversy is about documenting what TLS would look like if you needed pure MLKEM.
The "NSA" stuff is just more innuendo.
Apart from bytes on the wire, MLKEM is not in fact "much more expensive" than 25519. Also, you have the maturity argument wrong; it's not a concern that risks/flaws will be discovered "in the standards".
There is no hard evidence that it was backdoored.
https://archive.nytimes.com/bits.blogs.nytimes.com/2013/09/1...
This is not true. I can build Signal from source from GitHub, and use Signal-the-service with the client (which did not come from Signal, but GitHub/my compiler).
Many cryptosystems are like this. In any case, if you are getting something from the App Store, you can get it once and disable autoupdates, which prevents the service provider (presuming they are the same as the people who published the app) from backdooring you at some point in the future. Alternately, even with updates, unless Apple is colluding with them to serve only you* a specific backdoored app, you can at least be reasonably confident that it's not specifically backdooring only you* in an undetectable fashion.
Sure, but can you find an NSA-designed backdoor in the source code?
> you can get it once and disable autoupdates
Try doing that with Signal, and you'll be unable to connect to the main network in just a few days because you get out of sync. Also, what do you do if there's a high severity CVE on the program? You still don't update or you re-audit all the new code?
What you describe may be possible for an intelligence agency, but completely out of reach for an individual.
> unless Apple is colluding with them
Given the most likely adversary is the US intelligence, it's absolutely not far fetched to assume that in your threat model.
> you can at least be reasonably confident that it's not specifically backdooring only you
That's not really reassuring…
> Sure, but can you find an NSA-designed backdoor in the source code?
You're moving the goalpost. They were responding to the claim suggesting it's impossible to get non-Signal provided signal.
>> you can get it once and disable autoupdates
> Try doing that with Signal, and you'll be unable to connect to the main network in just a few days because you get out of sync.
That's demonstrably false. On one of my idle/backup phones I'm using Signal 8.8.2, released in April 2026, almost 3 full months ago. It can not only connect to the network but everything works, with every contact.
You might think of the official Signal client expiration, but that's client side (meaning that you can compile and use the version that doesn't have it) and..... 90 days, not "a few".
I don't have a concrete number for the server side of enforcement though (minimumVersions seems to be populated at start time, with the defaults not committed to the repo). It's not entirely unreasonable to assume that the lowest official supported version is the one that introduced the concept of usernames, and the only meaningful capability test is SPQR.
> Also, what do you do if there's a high severity CVE on the program? You still don't update or you re-audit all the new code?
I think disabling auto update was shown as a possible strategy against a silent, targeted auto update. Not a way to remain protected against the general CVEs.
Non sequitur.
That was never my claim. The claim is that you cannot protect youself from Signal being malicious if Signal is the maker of the software. Compiling the software yourself doesn't help against the kind of adversary in the threat model.
> That's demonstrably false. On one of my idle/backup phones I'm using Signal 8.8.2, released in April 2026, almost 3 full months ago. It can not only connect to the network but everything works, with every contact.
Lucky you, you only need to fully audit the codebase every 3 months.
I'm using the Signal apk directly so I'm painfully aware of the frequency of the breakages.
> I think disabling auto update was shown as a possible strategy against a silent, targeted auto update. Not a way to remain protected against the general CVEs.
I don't think you understand my point. I'm not talking about the CVE being exploited about you. The CVE will just push you to download the compromised update, breaking your “security through lack of update” policy.
Backdoors can be well-hidden, but this is the kind of software that people look for vulnerabilities in [1]. This gives a lot less flexibility to any potential backdoor.
Additionally, someone could decide to write a new Signal client from scratch, designed to be compatible with the original server. Such a client would probably be less secure overall, at least at first, but it couldn't have a backdoor inserted by Signal developers, since it wouldn't contain any Signal code. Since the original client also supports end-to-end encryption, a new client can be compatible with old clients.
[1] https://community.signalusers.org/t/overview-of-third-party-...
This. The author is dismissing the whole web-based cryptography, or any end-to-end cryptography for that matter, on the basis of a one-dimension analysis.
Not necessarily. I push for e2ee everywhere I can for a completely different reason: when (not if, when) we get breached, we cannot leak sensitive data we don't have.
(Just as an example, at my current employer, out of 8 people in the company, there are 8 people who have credentials that permit at least some db access in one way or the other (anyone that has at least one of this job: customer support, account managers, data scientists, devs). There are only 3 people who can push changes to production, and you need 2 out of 3 to do so, and all of three get paged when a release happen. Of course it's not infallible, but the probability of the app being corrupted is orders of magnitude lower than the DB being breached, by the mere scale of the number of people who have access to these things. And again, you only need transient access to the data to leak it, when you need persistent access to do damage by corrupting the code).
That's not completely true. If I can control when (and if!) the software updates and if there is some kind of vetting process to verify that the version I'm currently running does not contain a backdoor, I can treat it like a third party with respect to the server.
I agree with you though that most current software that are made to auto-update at any time without any oversight do not fall under this umbrella. Web apps definitely don't fall under it.
This would be extremely difficult, I would say impossible from a practical standpoint.
You don't need E2E for that, using https/TLS for transport and servers hosted in the US would be enough.
Data breaches happen literally every day.
The definition is quite clear. It does not apply when the implementation is not distributed by the same entity that creates it for example. There are other related issues but the message here is that web based cryptography has a particular weakness when it comes to things like end to end encrypted messaging.
How can you be sure that the entity distributing the software didn't backdoor it?
> the message here is that web based cryptography has a particular weakness when it comes to things like end to end encrypted messaging
There's literally no substance about that claim in TFA.
If the software is open source and you only install new versions after their source code has been audited, you should be ok.
But with repro builds and system transparency, hiding backdoors is impractical.
On other hand its quite natural, security is not really getting you direct revenue so business is least motivated in investing it or say continuously investing in it. The ones that do are doing partial lip service for most part.
Framing this in terms of governmental espionage is nonsensical. Using e2e from a US-based entity makes you completely sure that the US government is spying on you, because they assert direct control over the software you are running. There is no venue to seek justice against an unlawful contract if the government is in on it.
You should instead take a step back into reality and consider data misuse by normal non-government actors. Facebook claiming e2e encryption is a contractual matter, that you can litigate. That's where the actual protection is. That's also why real business demands not "unbreakable encryption" but reliable marker for access and tampering. It is much more useful to have a record of who accessed the data, than a claim that it's impossible.
Well that’s just wrong, just make it open source and do reproducible builds
1. Aren't E2EE systems designed to prevent decryption of content already created in the past sitting on the vendor's servers? Yes, the vendor could go rogue, but, assuming they currently have implemented E2EE right, it means any change to the client can only compromise content created in the future from that point onward, no? So why is the article implying Apple could have provided a back-doored iOS to bypass the encryption for existing content?
2. I also don't find the argument that E2EE is only a legal trick fully convincing. There are several other incentives for a vendor to implement it apart from avoiding legal issues: preventing insider abuse, reducing liability, improving customer trust, and resisting mass surveillance
These are real engineering motivations. The threat model is not: "Protect you if <vendor> becomes actively malicious tomorrow." Its more like "Protect messages stored on <Vendor>'s servers from attackers, employees, hackers, routine legal requests, and passive surveillance."
* preventing insider abuse * reducing liability * improving customer trust * resisting mass surveillance
Isn't this conflating encryption with trust? Of course whoever claims to encrypt your data needs to be trustworthy, and whether they actually are is another matter, but If my app allows you to generate a client side key, export it and use it to encrypt data client side and we only get the encrypted data, that is verifiably valid encryption.
I could be malicious and also send a copy of your actual plaintext to the server as well, but that is trivial to check (unless I'm being targeted and I am the only user that gets the malicious code, still, I can check). It's a risky proposition for an organization with vested interest in being seen as pro privacy.
But I get it, different conversation if the government coerces you, and the outcome depends on your bank account and ability to handle pressure.
Also no OS integrated system that does this for you automatically / conveniently has ever existed that was widely adopted because that application would have the ability to read all of your private communication, and impossible to install on an uncracked phone.
Still it would take literally minutes to vibe code an app that sits in front of a WhatsApp client and automatically handles these things. Maybe the future is just to write it yourself (not the security) so you can trust it and it’s convenient.
A sophisticated actor might as well also control the application that ends up on my device. It does not have to be the same delivery mechanism as long as I did not write it myself.
So all cryptography is snake oil?
___
I mean I kinda sorta get the point and there would be some merit to discuss there, but the weird framing makes that very hard to do.
Of course it's easier to break web e2ee if you are for example cloudflare compared with someone also having to compromise the Debian repos.
But that's not what snake oil means.
The xz utils hack got slurped up into sid before it was discovered by a researcher's performance regression in ssh. IIRC the hacked test file didn't even need to be added to the upstream source tree because Debian was blithely downloading release tarballs from Github. No evil Debian maintainer needed.
It's funny that when speculating about Debian's security you forget an actual state-level attack that got code into sid, but when speculating about Signal's insecurity in another thread you're quite happy to imagine potential state-level attacks.
Yes, but it's a whole lot of extra steps spread across multiple independent parties, each of them adds large delays to the actions and increasing the chance that it is discovered long before it ends up on the users machine.
When you hack GPG it will take years before it trickles down into every Linux distribution, especially LTS releases. And ideally, you want an encryption protocol, not one app, thus you have some people running GPG, some running Sequoia PGP and some running OpenPGP.js. If somebody fiddles with the encryption, different clients won't be able to decode the messages anymore and it will be clear pretty quickly that something is wrong.
Meanwhile on the Web or smartphones, you remove or backdoor the encryption, everybody gets auto updated to the latest version and nobody will know that something went wrong.
This reminds me Telegram, which promises to be secure, but requires giving it my phone number, which is the most insecure thing one can do.
for this to work in practice it needs to be paired with reproducible builds, open source and either p2p or server choice (use signal.mydomain.net instead of signal.org). but these are all things that already exist and none of them is really hard to set up. the harder problem is distributing community block lists of bad package versions but that can be done with atproto or simple ublock style filter files.
i think the real bottleneck for adoption is that the only browser with built in ipfs support is brave, the one thats full of crypto ads and affiliate link fraud. i dont know if firefox would ever take it up or we need to build a brand new browser. or find a way to do it one layer down with a system service.
This means that, in practice, iMessage is not e2ee.
Before you say "But what about Advanced Data Protection that enables e2ee for iCloud Backup?" - virtually nobody has this on, Apple prohibits you from turning it on in the UK, and even if you enable it - the people you iMessage with don't, so your conversations are in their backups. This means that if either endpoint of the iMessage conversation is in the UK, and both parties have iCloud Backup enabled (the default), then your iMessages are not e2ee as a non-endpoint has an escrowed copy of the plaintext or keys.
This effect was seen in the Apple vs FBI incident described in the article. The public perception of Apple as a brave defender of their user's privacy was greatly increased due to that dispute. For all we know, the FBI was in on the conspiracy. In return they might receive the fruits of such surveillance with the only limitation that they would have to disguise the source with parallel construction.
[1] https://en.wikipedia.org/wiki/Crypto_AG
E2E makes it less likely that your information will get hacked and reduces the risk that employees will access your information.
The reality is that these security claims are generally subject to internal audit and would need company wide collusion and the risk of a whistleblower or disgruntled former employee if they were violated provides some level of protection that a large tech company offering of e2e doesn’t mean some level of benefit from the user compared to perfect encryption security.
The host could inject malicious JavaScript from the host or change libraries but I feel like this is an avoidable problem because it can be audited much more easily than expecting users to audit JavaScript every time. People could even build known, trusted, web frontends. So I think there are mitigations if not ways to assure the browser is running trusted code.
(Said tongue-in-cheek, I don’t know if there’s other comparable systems out there)
It always comes down to who the alternative party to trust would be. In the fictional dialogue, the alternative appears to be to not send the message. Which may or may not be the better option than to give Eve eavesdropping capabilities.