Hacker News Reader

Comment by tptacek | Hacker News Reader

Section 8.2 seems to talk about how you send plaintext email via SMTP to users who aren't using Skiff.

But that's not what I'm talking about with respect to end-to-end encryption. The white paper refers repeatedly to "browser" users. Your server can feed arbitrary Javascript to browsers and subvert encryption in a variety of ways, can't it?

I'm still not clear why you designed a new, simplistic cryptosystem at all here; can't you do everything you're trying to do here on top of Matrix? Again: the cryptography promises you're making only work between users of your system.

View on HN

> Your server can feed arbitrary Javascript to browsers and subvert encryption in a variety of ways, can't it?

That's how literally any website works. How do you encrypt in the browser if the server doesn't send JavaScript to encrypt data? You also trust Signal not to issue an update that sends data in plaintext over the network. Unless you're building an app from source, you implicitly trust the developer to some extent.

Signal doesn’t have a web client. Most of the stores it is distributed through have fairly strong resistance to compel orders.

Yep, but Signal is still a potential adversary, and could roll out a backdoor.

A couple of things that are easier in a web-delivered tool is deliver a backdoor to a user or group of users (which Skiff can track), or deliver a backdoor over a particular window of time across many users to decrease the chance of detection.

I know Skiff uses IPFS in some of parts of their solutions, and there's something they could do with that for the first -- essentially making visiting a particular version of the code part of how it is accessed, but there's some real UI challenges, which maybe they're looking into (it's been a while since I checked them out: they have some great UX in other parts of their suite).

The other tactic I've seen is to bundle the page into a browser extension, which moves you closer to Signal's status.

Signal's servers can't backdoor the Signal Client. From what I understand of Skiff, Skiff's servers are the client.

Not directly. They would have to roll out an update with the backdoor to the App Store. But as a user I’d be none the wiser.

I wish there was some way on iOS to prove that some particular version of an app was built from a certain git hash. That way these sort of attacks would be easier to detect.

That would be ideal. F-Droid on android can do this: https://f-droid.org/en/docs/Reproducible_Builds/

However there is still one advantage of even appstor: They have to push the backdoored version to everyone (or large set of users). So that drastically increases risk of being caught. Website under their control can backdoor one specific user or even just one session, making detection harder.

> So that drastically increases risk of being caught.

Only if someone out there is extracting, decompiling and auditing each version of the Signal iOS app in the app store. But I doubt anyone is doing this. If a backdoor is ever snuck into the signal ios app for a few users for a few weeks, I highly doubt anybody would notice.

xkcd://386

Of course someone is doing this. I’m not sure they are the kind to tell it to the world, though.

Not just someONE, whole bunch of researchers and automated scans.

You can change remote configuration flags post-release (e.g. to enable diagnostics).

Even “secure” softwares like Google Chrome can capture your whole browsing history if they suddenly decide to enable a flag on your IP address. No need for conspiracy or update, though Chrome is considered perfectly secure.

In Android you can also distribute updates to specific e-mail addresses, which is very convenient.

> In Android you can also distribute updates to specific e-mail addresses, which is very convenient.

Yes but this requires the user to opt-in, you can't do it silently:

> After clicking the opt-in link, your testers will get an explanation of what it means to be a tester and a link to opt in. Each tester needs to opt in using the link.

Source: https://support.google.com/googleplay/android-developer/answ...

As for the Chrome thing, I'm a Firefox user but I would be surprised if it shipped with the option to remotely upload whole history without user's knowledge or consent, do you have a source to back that up?

You can disable updates. As long as the servers (and the OS) don’t break compatibility, you can continue to use the same old likely-non-backdoored version.

Not for Signal. I was periodically prompted to update with the warning that I would lose access to signal if I didn't. Then I upgraded and they robbed me of SMS features. I would have been happy to remain on a previous version.

At least when downloading from the Play store, Google requires the application be updated every 90 days or so. I've run into this multiple times where the application ceases to function after the 90 day window until I go update.

Group invite links in Signal work similarly to Skiff: they reveal a key to the client side js so it can either present users with a Signal download link or open a URI, depending on whether they already have Signal or not. (Signal chooses not to use a URI in this case because it would break for anyone without Signal.)

So if a Signal group is using group invite links they have the same problem Skiff does.

What I would like to exist is something like Subresource Integrity [1] but for a URL itself, so that you could include a hash in a URL and let the browser warn you if the page source doesn't match the hash.

1. https://developer.mozilla.org/en-US/docs/Web/Security/Subres...

> How do you encrypt in the browser if the server doesn't send JavaScript to encrypt data?

Meta has done some work along with Cloudflare on this for WhatsApp Web, specifically. In general, JS crypto is always going to be suspect if the threat model involves distrusting the server (like in e2ee protocols like Signal).

JS crypto functions now interface with browser crypto functions for the last decade or so. https://developer.mozilla.org/en-US/docs/Web/API/Crypto

What difference does this make to the thread model in the previous comment?

Before this basic cryptography was downloaded via JS files which yields no security and gave web cryptography a bad reputation. That is not true now.

Huh? It is very concerning to hear this from a founder. It is the same exact level of security as it is executing the same code, just at a different level. Really does not matter what crypto lib you are using if at the end of the day the surrounding code dictates all the security.

Regardless of all this, your open source crypto library doesn't even use the Web Crypto APIs at all, but rather the dreaded js based crypto you are badmouthing (tweetnacl, stablelib)

That's just false. Downloading crypto libraries over the web plagued Javascript crypto for years. We use tweetnacl, stablelib, and webcrypto - and tweetnacl also uses webcrypto!

From your source, chacha20 is used, which is literally not supported by web crypto apis. Again, this makes 0 difference as higher level implementation makes or breaks your crypto.

There's literally nothing stopping the page from simply not encrypting the data.

I built something that solves this problem. It uses an open source program installed on the client to encrypt and store data in a way that doesn't allow the website to access the plaintext data. https://redact.ws

It would be far more illogical to warp Matrix into an end-to-end encrypted calendar, note-taking product, file storage product, and email platform.

Why? You'd use Matrix to distribute keys for encrypted blobs. That's how systems like these work anyways.

What do you think of Lavabit? I think they operated in the same way, but the US government forced them out of business for refusing to hand over their TLS keys to allow the US to spy on Snowden.

https://en.wikipedia.org/wiki/Lavabit

Lavabit is the one that used user passwords to encrypt the messages, thus ensuring that they had access to all the necessary secrets to decrypt user messages any time the user was viewing them?

And that had complied previously with US government subpoenas to provide metadata and data for users?

Isn't that what you're Skiff is doing too? It seems like it's just the Lavabit design with a some 2010 cryptography layered on top.

No. Lavabit had fundamental flaws where passwords were sent to the server so anyone who could decrypt the HTTPS traffic could basically access the content [1].

Skiff's password mechanism actually solves this flaw cryptographically using known, established primitives.

We use argon2id to take password and turn it into two cryptographic keys. One key is used for an SRP scheme to prove you have the password in a signing flow that bootstraps session management. The other key is actually the data encryption key. These keys are never sent to Skiff's servers and this generation happens all in the browser.

Lavabit really failed fundamentally in having actual end to end encryption because the password was sent to the server.

[1] https://arstechnica.com/information-technology/2013/11/op-ed....

That's all moot though, since your server has access to plaintext emails when sending and recieving (99.99999% of email addresses would be outside skiff), which completely subverts the whole point of encryption. A vulnerability on the server could leak all user emails, without needing their keys.... This is a solution theoretically only as strong as encryption at rest.

>And that had complied previously with US government subpoenas to provide metadata and data for users?

Interesting. Link? Are you talking about this article? https://www.forbes.com/sites/kashmirhill/2013/08/09/lavabits...

It seems to be talking about metadata, not data.

See below, Lavabit not a good comparison as it was not end-to-end encrypted. Also read https://arstechnica.com/information-technology/2013/11/op-ed...

Thanks for the information.

One similarity between Lavabit and Skiff is that tptacek calls them both non-end-to-end encrypted.

It's very simple. One of them had access to user's private keys (Lavabit).

One never has access to user private keys (Skiff).

I don't understand how you don't have exactly the same access they did. I feel like I've invested a fair bit of time to understanding how this stuff works, and the story you're telling doesn't make sense. What am I missing?

What amilich said is correct. However, what he is leaving out is that both have access to unencrypted email at send and recieve time, so you are taking Skiffs word that they dont log emails - since you have to trust the server this is not e2ee.

Reply on news.ycombinator.com