Hacker News Reader

You have to send emails through third parties or people won't get them, because you are also always sending them to third parties who host the recipients email and manage their spam. It might be a good reason not to send magic links but here we are talking about a tertiary confirmation, so its useless on its own right?

The link in the email could be a direct link to Mercury's website, rather than one that passes through a third-party HTTP redirect service.

Authentication tokens (even tertiary ones) usually are supposed to have pretty strong secrecy guarantees. I've done multiple security audits for SOC, PCI, HIPAA, etc., and in every case the auditors would have balked if I told them signin tokens were being unnecessarily logged by a third-party service.

(Also: I strongly disagree that the only way to get reliable delivery is via a third-party email service, especially at Mercury's scale, but that's a digression from the topic at hand.)

Oh good find, the link going through Mailgun as a redirect is a recent regression. We have a PR to fix that going live soon.

That said, our security team and I agree there is no security issue here. Mailgun already can see the text of the emails we send.

How is there no security issue here? Email is not secure and it's even less so when you are sending it via a 3rd party. If this were a photo site or something that would not be a big deal but we're talking about a bank. SMS is not much better. Like somebody said elsewhere in the thread, you should allow people to opt out of insecure third-factor verifications since they are just an annoyance and are ultimately security theater.

It's not security theater. He explained above how this is used to defeat a specific phishing attack that they've actually seen in the wild. There are other, different threat vectors (e.g. compromise of the mail server) that it doesn't prevent. But that doesn't make it theater. as it does provide other value.

What does it stop? You already did a 2FA at this point. If an attacker has my 2FA he most likely already has my email so the 'value' being provided is at the cost of more complexity for the user. If this adds value then why not also do an SMS as well to be really, really sure that the user is legit? That would add even more value.

And again, I wasn't saying that you can't do all of this nonsense, but users who see it as nonsense should be able to turn it off.

Why would the attacker having your Mercury TOTP mean they most likely have access to your email?

Because my TOTPs are all stored in the same device and in my imaginary scenario they have that.

Again, see the post by MaxGabriel at https://news.ycombinator.com/item?id=42629109 where he explains how this measure actually defeated that particular pihishing/MITM attack.

The attack wasn't that the attacker has my second factor, the attack was that the attacker tricked me into verifying a single login/transaction using my two factors, on their behalf.

They probably judged that the inconvenience of the verification email affects few enough users that it is worth it. Most users don't switch IP addresses very often. And those that do, probably don't all clear their cookies after every session.

Adding SMS in addition to email would be obviously useless, as you point out.

The emails in question are a third factor, not a magic login link.

Even if they were, almost all email goes through third parties which are trusted implicitly. That's not great, but email is the only federated system in existence capable of implementing this type of decentralized login at scale.

Maybe someday we'll be able to use something like Matrix, Fediverse OAuth, or ATProto OAuth instead, but those are all a ways off.

> passes through a third-party HTTP redirect service

The vendor might not be the only party to use an HTTP redirect service too! My email goes through a security screen by $EMPLOYER, which also rewrites links to get processed through their redirect service. Sure, it's for company-approved reasons, but it's still another party that has access to the login token.

What would be a more secure (yet reliable) method for email delivery for such emails?

Make the link in the email https://mercury.com/something instead of https://mailgun.com/something (which then redirects to https://mercury.com/something). Or (in addition to, or instead of, a hyperlink) provide a 6-10 digit numeric or alphanumeric code that could be copied out of the email message into a form field on the signin screen.

> 6-10 digit numeric or alphanumeric code that could be copied out of the email message into a form field on the signin screen.

To be clear this is what we're trying to avoid. An easily typeable code like that can be typed into a phisher's website.

How about giving me a setting to disable device verification: "I know how to type mercury.com into the URL bar and accept all risk of getting phished."

I appreciate you guys are trying to protect people, but no other financial institution I deal with requires this level of annoyance, and at some point I'd rather switch to a less "secure," but more usable service.

(I put secure in scare quotes, because some suggestions, like trading true 2FA, where I have two separate secrets on two separate devices, for a single WebAuthn factor, are actually accomplishing the opposite, at least for those of us who don't click links in emails and don't use ads on Google for navigation.)

Edit to add: or maybe save the third factor for suspicious activity, such as "new device adding a new payee," rather than every signin. It's been months since I onboarded a new vendor, and I'd be OK with only having to do the cut-and-paste-the-link dance a couple of times a year, rather than every single time I want to check my balance.

My understanding (as CEO of a startup using Mailgun for magic links) is that you're seeing mailgun in the URL because they have click tracking enabled — which, to be fair, is not super useful in the case of verification emails.

They could use a custom subdomain for this click tracking and "hide" the mailgun url from you, but we're finding that for some reason Mailgun doesn't just use a let's encrypt certificate, so some users will complain that the tracking links are "http" (and trigger a browser warning when clicked).

Anyway, even with click tracking disabled and links going straight to mercury.com, the security issue would remain the exact same (since Mailgun logs all outgoing email anyway).

But my understanding is that the contents of that email and its link do not provide "login" capability but "verification" capability. As such, a Mailgun employee accessing your data, or an attacker accessing your Mailgun logs, would only be able to "verify" a login that they had already initiated with your password AND your OTP —which means that's effectively a third hurdle for an attacker to breach, not a one-step jump into your account.

At the very least, you can be creative with workarounds for such issues. A bookmarklet can be convenient.

    javascript:void(window.location.href = window.prompt().replace(/\\\n\s*/g, ''));

I set Firefox to delete cookies at shutdown, and also an add-on called Cookie AutoDelete, but they both have an option to whitelist a site.

> I wasn't aware that WebAuthn didn't have this requirement. I prefer TOTP because I actually like having a second factor in addition to a credential stored on my computer's hard drive (whether a password or a private key in my password manager), but I might be willing to reduce my security posture to get rid of this annoyance.

I've seen passkeys support something like what you're after. The browser will produce a QR code you scan with your phone, and then you authenticate with the passkey via the phone, which then authorizes the original browser.

I'm not absolutely certain that this is part of the spec or how it actually works. I'd like to know. It solves a couple different usability issues.

You could always use something like a Yubikey.

> You could always use something like a Yubikey.

This is the option I prefer, but only on sites that allow me to enroll more than one device (primary, and backup for if the primary gets lost or damaged). AFAICT, Mercury only allows a single security key.

I have an encrypted offline backup of my TOTP codes, so if I drop my phone on the ground, I don't get locked out of all my accounts. I keep this separate from the encrypted offline backup of the password manager on my computer, and as far as I know, neither has ever been uploaded to anyone else's "cloud." Malware would have to compromise two completely separate platforms to get into my accounts, rather than just iCloud or whatever credentials.

I understand the desire for phish-proof credentials, but—given that I don't click links in emails—my personal threat model ranks a compromised device (via attack against a cloud service provider, or software supply chain attack against a vendor with permission to "auto-update," or whatever) much higher likelihood than me personally falling victim to phishing. I readily admit that's not true for everyone.

> AFAICT, Mercury only allows a single security key.

We allow multiple security keys. You can add more here: https://app.mercury.com/settings/security

Oh, nice! This wasn't obvious from the help text. Maybe add it to the FAQ on the "Adding security keys" sidebar?

> my personal threat model ranks a compromised device ... much higher likelihood than me personally falling victim to phishing

I completely understand that. I'd actually be interested in reading anything practical you might have on that topic if you don't mind. I asked some experts who gave a talk on supply chain security last year ... they didn't have a lot of positive things to say. Developing software feels like playing with fire.

It feels unstoppable, which is why the best I can do is try to mitigate its impact. Some mitigations that come to mind:

The development environment where I'm downloading random libraries is on a completely separate physical machine than my primary computer. I generally spin up a short-lived container for each new coding project, that gets deleted after the resulting code I produce is uploaded somewhere. This is completely separate from the work-supplied machine where I hack on my employer's code.

On my primary computer, my web browser runs in an ephemeral container that resets itself each time I shut it down. My password manager runs in a different, isolated, container. Zoom runs in a different, also isolated, container. And so on.

Wherever possible, I avoid letting my computer automatically sync with cloud services or my phone. If one is compromised, this avoids spreading the contagion. It also limits the amount of data that can be exfiltrated from any single device. Almost all of the persistent data I care about is in Git (I use git-annex for file sync), so there's an audit trail of changes.

My SSH and GPG keys are stored on a hardware key so they can't be easily copied. I set my Yubikey to require a touch each time I authenticate, so my ssh-agent isn't forwarding authentication without a physical action on my part. I cover my webcam when not in use and use an external microphone that requires turning on a preamp.

I try to host my own services using open source tools, rather than trust random SaaS vendors. Each internet-facing service runs in a dedicated container, isolated from the others. IoT devices each get their own VLAN. Most containers and VLANs have firewall rules that only allow outbound connections to whitelisted hosts. Where that's not possible due to the nature of the service (such as with email), I have alerting rules that notify me when they connect somewhere new. That's a "page" level notification if the new connection geolocates to China or Russia.

I take an old laptop with me when traveling, that gets wiped after the trip if I had to cross a border or leave it in a hotel safe.

I have good, frequent backups, on multiple media in multiple offline locations, that are tested regularly, so it's not the end of the world if I have to re-install a compromised device.

> The development environment where I'm downloading random libraries is on a completely separate physical machine than my primary computer. I generally spin up a short-lived container for each new coding project, that gets deleted after the resulting code I produce is uploaded somewhere. This is completely separate from the work-supplied machine where I hack on my employer's code.

Something like VS Code remote dev with a container per project? Just plain docker/podman for containers?

> On my primary computer, my web browser runs in an ephemeral container that resets itself each time I shut it down. My password manager runs in a different, isolated, container. Zoom runs in a different, also isolated, container. And so on.

Qubes, or something else? I've been looking at switching to Linux for a while, but Apple Silicon being as good as it is has made making that leap extremely difficult.

Mostly Linux with systemd-nspawn, also some Kubernetes, plus the occasional full VM. (If I were setting this up from scratch, I'd probably try to figure out how to run my desktop as 100% Kubernetes, using something like k3s, but I don't know how practical things like GPU access or Waypipe forwarding would be via that method.)

I live inside Emacs for most things except browsing the web, either separate instances via SSH, or using TRAMP mode.

If you switch to Linux, I highly recommend configuring your browser with a fake Windows or MacOS user agent string. Our Cloudflare overlords really, really hate Linux users and it sucks to continually get stuck in endless CAPTCHAs. (And doing so probably doesn't hurt fighting against platform-specific attacks, either.)

Is there a reason that TOTP cannot be used as a second factor when using Passkeys?

Not sure why we suddenly went from 2 factors (password + TOTP) to 1 factor (passkey), even if passkeys themselves are better.

TOTP should at least be an option for the users.

Pretty sure that is eMacs formatting, not the email itself? Can you kill-copy the URL?

So you are intentionally crippling your browser and ability to access email (you need to ssh to another computer and access via terminal). You also aren’t able to handle emacs wrapping of long lines. And you are complaining that the security in place to prevent stolen credentials is “inconveniencing you”.