I wrote to them pointing out the issue and apparently it was put on their backlog for the following year. I think I received an update about it a couple of years later.
Asking for random letters is a poor man's version of a zero knowledge proof. It's only marginally more secure than asking for the full password (if you want to break into an account you need to observe someone entering the password a few times instead of just once), at the cost of a very crappy user experience.
The key difference between storing plaintext passwords and hashed passwords is that with hashed passwords, you can only compromise users who log in during the interval when the system is compromised, and even then, only for worse compromises (eg if all the attacker gets is arbitrary disk read, compromising hashed passwords would be impossible in most scenarios.)
Why only then?
There is a standard for it, called TLS-SRP, but unfortunately no browser supports it.
Though, my bank uses a dual-password setup where the first password is required in its entirety and a few letters of the second password are required. I suspect this is stored by using a (hopefully memory-hard) salted key derivation function/password hash function (such as Argon2 or scrypt) to derive an encryption key from the first password, which is then used to decrypt the second password.
At least, that's the way I'd design the system to minimize consequences of database compromise. Users are often not creative and often use related passwords. You'd want all of the stored information about the second password to provide clues to the first password only if the attacker had already cracked the first password. Hashing all of the permutations really blows up storage space and DB I/O.
Ok, yes. Encryption could work, but encrypted passwords aren’t much better than plain text ones.
I think a lot of implementations don't do anything clever and just have a fixed set of combinations they use each time, so the number of stored hashes is smaller.
With Shamir's Secret Sharing, you start out with your secret expressed as a finite field element, and pick N-1 polynomial coefficients uniformly randomly from your finite field, and give M different participants values of the polynomial evaluated at different points. Information theory shows that having at most N-1 points on the polynomial still leaks no information about F(0). It's as secure as a one-time pad, and in the degenerate case of N=2 over a Galois field, is in fact a XOR-based one-time-pad. The amount of entropy added at generation time is greater than the amount of entropy in the secret.
In this partial password case, you'd want the "shared secret" to be a single value, and construct a polynomial such that, say 4 field elements, would always be sufficient to recover the same "secret". You're running the algorithm the wrong way, and unless you're forcing a password on the person, those letters are actually low-entropy instead of maximal entropy field points as in Shamir Secret Sharing.
I really don't see a way to use polynomial interpolation in this way without leaking entropy. I thought about using something similar to Schnorr threshold signatures[0] to solve the information leakage, but then you end up storing one elliptic curve public key per password letter, and the private key for each public key is a single letter, so trivial to bruit-force.
But yeah, that's infuriating.
It plugs in as an HID so that passwords appear to have been typed in, not PW managed.
Useful little thing, it seems too good to be true.
Edit: I'm also a bit uneasy about securing access to all by accounts by one single PIN that can be used to unlock a physical device that is easy for someone to steal unnoticed.
But it turns out that keeping that 1 login and password in your head and trusting the rest to your password service (especially if you let it make random passwords) is way more secure than what people do if they have to try to keep them all in their head.
So it's about relative security. Both sides are correct, in their narrow views of the situation.
But it depends on your threat model and attack surface area/shape. A strong password on a post-it on your monitor is stupidly insecure locally, but far more than just having a weak password if you consider only non-local (to the user) attacks (neither that site hacker in far-off Hackyoustan nor his pet bots can see the post-it, but could try brute force a short easy to remember password).
But you're right, browsers would have to become picky on saved credit card auto populate to force a change.
Unlikely. The password managers would just lose users.
I've even caught myself being annoyed at 1Password when it doesn't work while the cause is obviously the website doing some stupid thing.
I wonder how 1Password works on this problem though, do they just accept the crap and keep working on their guessing magic or do they also invest in standardization?
Hopefully hardware tokens can help us get rid of passwords soon.
Now in this situation, it looks like Substack's UI is to blame. The secret form field appears to be heavily modified from its original form state, probably masking the blue background that 1Password attempted to give it.
1Password is in a difficult situation here. If the autofill doesn't work (because the developer of the website they are filling, uses non-standard naming), then users will blame 1Password for "not working", not their browser or the website developer. 1Password therefore has to make it work with as many forms as possible around the web, most of which are built by lazy developers trying to throw up a website during a hackathon, or from a developer that learned to code from a 1 hour youtube video. It's a tall order, from their perspective, to try to make this work with as many forms as possible, while having the fewest (ideally none) false positives.
The one thing I can recommend for anyone who uses any password manager, is that before you click submit on a form that you used auto-fill for, you always look over the entire form again. Now in OP's case, I could see how they could have easily missed the extra amount (again, Bad UI substack), but this is general rule that anyone with a password manager should follow. Double check everything before you submit.
This is a market problem. All but the last password manager to do this would go out of business, due to reduced utility vs alternatives.
This seems to be particularly common in the web development world, even in the fundamental tech upon which the web is built. Missing the closing tags in your HTML? The browser will just guess where it should be for you. Missing a semicolon in your Javascript? The interpreter will just guess where it should be for you. Invalid CSS? The browser will just silently ignore it. Illegal comments in your JSON? Most parsers will just assume you meant to ignore those lines.
Other examples are problems with lock files and file versioning (because programs tried to roll their own when the operating system didn't provide them), and the complexity of parsing "HTML soup" and emails with all kinds of bizarre invalid syntaxes.
I can't offer a general solution, but if password managers simply refused to autofill to any field other than the one with the matching standard autocomplete attribute, web developers might start doing the right thing. (Do Safari, Chrome and Edge already do this? Only they have the clout to make it happen.) The user could still fill out a text box lacking the standard autocomplete attribute by right-clicking and manually selecting the correct field. Password managers should also get cheaper because their vendors would not need armies of developers adding workarounds for popular sites.
Somehow this kind of nonsense has become culturally acceptable in the software industry. If the car industry worked this way you'd have to take your car back to the dealer once a month to be patched to take account of constantly changing fuel formulations. Standards exist for a reason.