Alice and Bob Sign Messages

Alice and Bob can send messages privately via a nosy postman, but how does Bob know that a message he receives is really from Alice, rather than from the postman pretending to be Alice?
If they’re using symmetric-key encryption, and Bob is sure that he was talking to Alice when they exchanged keys, then he already knows that the mail is from Alice – as only he and Alice have the keys that are used to encrypt and decrypt messages, so if Bob can decrypt the message, he knows that either he or Alice encrypted it. But that’s not always possible, especially if Alice and Bob haven’t met.
Alice’s shopping list is longer for signing messages than for encrypting them (and the cryptography to real world metaphors more strained). She buys some identical keys, and matching padlocks, some glue and a camera. The camera isn’t a great camera – funhouse mirror lens, bad instagram filters, 1970s era polaroid film – so if you take a photo of a message you can’t read the message from the photo. Bob also buys an identical camera.
sign1
Alice takes a photo of the message.
sign2
Then Alice glues the photo to one of her padlocks.
sign3
Alice sends the message, and the padlock-glued-to-photo to Bob.
sign4
Bob sees that the message claims to come from Alice, so he asks Alice for her key.
sign5
(If you’re paying attention, you’ll see a problem with this step…)
Bob uses Alice’s key to open the padlock. It opens (and, to keep things simple, breaks).
Bob then takes a photo of the message with his camera, and compares it with the one glued to the padlock. It’s identical.
sign6
Because Alice’s key opens the padlock, Bob knows that the padlock came from Alice. Because the photo is attached to the padlock, he knows that the attached photo came from Alice. And because the photo Bob took of the message is identical to the attached photo, Bob knows that the message came from Alice.
This is how real world public-key authentication is often done.

  • Alice generates a key pair, consisting of her public key (blue key) and private key (not shown here, but it’s how she acquires the blue padlock)
  • Alice takes a cryptographic hash of the message she wants to sign (the photo) using a commonly known hash algorithm (the funhouse cameras that both alice and bob have)
  • Alice encrypts the cryptographic hash with her private key, creating a signature for the message (the padlock with attached photo)
  • Alice sends the message and the signature to Bob (they can be sent separately, but the signature is almost always attached to the message to create a signed message)
  • Bob fetches Alice’s public key (the blue key)
  • Bob takes a cryptographic hash (photo) with the same algorithm that Alice used (identical funhouse camera)
  • Bob validates the signature against the hash (unlocks the blue padlock) with Alice’s public key

Some things to note:

  • The message was publicly visible – there was nothing to stop the postman reading it. You could use encryption to protect it, but that’s a separate process – encryption and signing are orthogonal. You can encrypt, or you can sign, or you can do both.
  • We inverted what the keys and padlocks mean compared to the last Alice and Bob story. Today the key is Alice’s public key, last time it was Bob’s private key.
  • The key pair for signing messages is generated in exactly the same way as the one used for encrypting messages. It’s very common to use the same key pair for both signing and encryption. When you send a signed, encrypted message to someone else you sign it with your private key and encrypt it with their public key. When your recipient receives the message they decrypt it with their private key and validate the signature with your public key.

How does DKIM use this to sign messages?
The algorithm (funhouse camera) that’s used to generate the hash is defined in the DKIM-Signature header by the c= and a= headers (there are several steps, but the final one is the cryptographic hash proper, typically SHA256). The hash (photo) is stored in the bh= field of the DKIM-Signature header, and the signature itself (padlock-with-photo) is stored in the b= field. The recipient uses the d= header to find out who signed the message, and retrieves the signers public key (blue key) via a DNS lookup. Then they use that public key to validate via RSA the signature (unlock the padlock). They then normalize and hash the message in the same way as the sender did, and confirm that the hashes match.
There’s a more detailed description at dkimcore.org and the formal specification is in [rfc 4871].
About that problem …
There’s a huge hole in the process as described. When Bob retrieves Alice’s public key, how does he know it wasn’t tampered with en-route? If an attacker, Mallory, has access to the communication channel he can send a message claiming to be from Alice but signed with Mallory’s private key. When Bob requests Alice’s public key Mallory can intercept that too, and replace it with Mallory’s public key. That signature will validate, and Bob will believe that the imposter message he received from Mallory really came from Alice.
There are several solutions to that. The simplest is to rely on a supposedly more secure channel to retrieve a public key with. That’s the approach DKIM uses to sign messages, relying on the integrity of the domain name system to publish the senders public key. That’s really not terribly secure, as there are lots of ways to compromise DNS. But it’s simple and it’s “good enough” for DKIM’s purpose – being able to bypass some spam or phishing filters by masquerading as someone else isn’t a particularly valuable target. That’ll get more secure as DNSSEC support is deployed more widely. (If your DNS isn’t protected by DNSSEC you should probably put it on your medium-term list of things to consider.)
Another approach is to use a web or hierarchy of trust. But that, and certificates, is another article.
 

Related Posts

DKIM replay attacks

Replay attacks on DKIM signed messages
When you receive an email validly signed with DKIM by example.com that might not mean that example.com sent the email to you, or that they even sent this email at all.
What it does tell you is that at some point in the past, example.com signed an email with exactly the same headers and body and sent it to someone. That’s often close enough to the same thing. But if that original recipient were to resend the email to you completely unchanged then the DKIM signature would still validate when you received it. That’s not a bug; it’s one of the design features of DKIM that it typically survives mail forwarding.
That original recipient could also forward the exact same email to a million of their closest friends, and the DKIM signature would validate at each of those million recipients ISPs. This is one form of a replay attack, and it isn’t something DKIM prevents.
DKIM doesn’t prevent replay, but does mitigate it
Completely eliminating replay attacks over SMTP is difficult – it’s inherently a store-and-forward protocol, so there’s no way to have the sender and recipient do any sort of handshake to ensure that a particular signature is only used once. It’s not unheard of for email to be delayed for days, and delays of hours aren’t unusual, so allowing a signature to be valid for only a few seconds after it’s sent won’t work. And the design requirement that DKIM signature survive forwarding means that it has to survive the final recipient’s email address not being the same as the email address the mail was originally sent to so you can’t include the envelope recipient in the signature.
So what does DKIM do to mitigate replay attacks? The answer to that is surprising – almost everything DKIM does is there to mitigate them. The DKIM signature depends on the body of the message, the subject line and the content of any other headers the sender chooses to include; changing any of that will invalidate the signature. That means that while anyone can grab a copy of an email sent by, for example, paypal and forward it on to someone else, if they modify the content at all it will no longer have paypal’s signature. So an attacker can’t just grab someone else’s signed email and replace it with modified content – and if they can’t do that, where’s the benefit to a spammer or phisher to replay a message?
But all that work is for naught if you allow the attacker to choose the content before you sign the message. There are several ways an attacker can do that, but one example that’s particularly relevant today is ESP trial accounts.
I’m stealing your reputation
If you allow anonymous signups for trial accounts that let a potential customer try out your system you’ll want to put very tight limits on how it can be used, so as to avoid spammers signing up and spamming through your servers. Maybe you’ll limit the number of email addresses the trial user can upload, or the number of emails they can send. At the most extreme you might even limit the trial account to sending mail solely to the trial users own (confirmed) email address.
But if an attacker can send even one piece of email they create through your trial account to themselves, and you sign that email, they can take it and send it to a million recipients – and it’ll still have your DKIM signature on it so it’ll use your reputation to avoid filters and end up in the recipients inbox. And then the recipients will report it as spam, and all that spam will be counted against the reputation associated with your DKIM identifier. If you share a DKIM identifier (“d=”) across all your customers that could cause all your customer mail to start being rejected or sent to the spam folder. (Even if you don’t it could still affect your delivery negatively, as spam filtering systems – both automated and human – sometimes aren’t entirely rational or predictable).
Spam that’s sent like this will be a little “off”, compared to legitimate email – the To: field won’t have the email address of the recipient, for instance, and there’ll be no personalization in the Subject or body of the message. It’s no worse than most spam, and it’s more than balanced out by being able to hijack someone else’s reputation.
So if you provide any way for unvetted non-customers to send email through your systems you should consider adding some DKIM limitations to the constraints you already have on that mail path. Not signing with DKIM at all avoids the problem altogether, but also means you can’t demonstrate your DKIM prowess to legitimate potential customers. You might want to sign with a DKIM d= domain that’s different to your production signatures, perhaps even a completely different top level domain to avoid any risk of confusion (but don’t try and hide that it’s your domain – that’s what spammers do).
Other operational mistakes
There are some grubby corners of the email and DKIM specs that sometimes interact to cause other holes that this sort of reputation hijacker can exploit. I’ll talk about header duplication tomorrow.

Read More

Gmail FBL update

Last week Gmail started contacting ESPs that signed up for their new FBL with more information on how to set up mailings to receive FBL emails.
One of the struggles some ESPs are having is the requirement for DKIM signing. Many of the bigger ESPs have clients that sign with their own domains. Gmail is telling these ESPs to insert a second DKIM signature to join the FBL.
There are a couple reasons this is not as simple or as doable as Gmail seems to think, and the challenges are technical as well as organizational.
The technical challenges are pretty simple. As of now, not all the bulk MTAs support multiple signatures. I’ve heard that multiple signatures are being tested by these MTA vendors, but they’re not in wide use. This makes it challenging for these ESPs to just turn on multiple signatures. For ESPs that are using open source software, there’s often a lot of customization in their signing infrastructure. Even if they have the capability to dual sign, if they’re not currently using that there is testing needed before turning it on.
None of the technical challenges are show stoppers, but they are certainly show delayers.
The organizational challenges are much more difficult to deal with. These are cases where the ESP customer doesn’t want the ESP to sign. The obvious situation is with large banks. They want everything in their infrastructure and headers pointing at the bank, not at their ESP. They don’t want to have that second signature in their email for multiple reasons. I can’t actually see an ESP effectively convincing the various stakeholders, including the marketing, security and legal staff, that allowing the ESP to inset a second signature is good practice. I’m not even sure it is good practice in those cases, except to get stats from Gmail.
Hopefully, Gmail will take feedback from the ESPs and change their FBL parameters to allow ESPs to get information about their customers who sign with their own domain.

Read More

A brief DMARC primer

DMARC stands for Domain-based Message Authentication, Reporting and Conformance. What DMARC does is allow domain owners to publish policy statements in DNS telling receiver domains what to do with messages that do not authenticate. In addition, DMARC introduces the concept of “domain alignment.” What this means is that the authentication has to be from the same domain (or a sub-domain) as the address in the header-from: line. The idea behind DMARC is that organizational owners can use SPF and DKIM authentication to authenticate their actual domain in the header-from line. This moves authentication from a important but behind the scenes technology out to an end user visible technology.

Read More