>Using CAA is a good idea in general, but would it help in this case? The attacker would just request the exact cert configuration that is permitted by CAA. Maybe this helps if you can strengthen one validation method?
Author of ACME-CAA (RFC 8657) here. ACME-CAA can mitigate this because you can put a unique identifier for your ACME account in the CAA record, so it is not possible for an attacker to do this unless they can get your ACME account private key (or coerce the ACME service). This assumes you have DNSSEC-secured nameservers, of course, otherwise DNS requests can potentially also be intercepted when queried by the CA.
See RFC 8657 for a full list of security caveats. The RFC is designed to be more readable than most.
Blogpost by me with more background: https://www.devever.net/~hl/acme-caa-live
Small update: I've just written a blogpost with my thoughts on the incident. https://www.devever.net/~hl/xmpp-incident
I usually brood on blogposts for days before publishing them since I care a lot about getting things accurate, but this is a bit of the moment, so here goes. Always happy to get feedback by email or IRC: https://www.devever.net/~hl/contact
Is it just me, or would a recurring RIPE Atlas measurement be a great way to detect fuckery like this?
https://atlas.ripe.net/
not at all, since there were no Internet-visible routing changes.
atlas would be good for detecting the time Pakistan announced 0/0 or whatever.
I think atlas allows tls/ssl probes, so that could still be used to track unexpected changes in certificates?
Wuuut, I missed they finally enabled it. Everyone who has infra MitM in their threat-model should look into and enable this.
Also, thank you for your work as well as popping in here.
Am I overstating it by seeing it as one of the most important milestones for Internet security in general over the past few years?
Do you have a take on DLT-based systems and if this is something that is (or could be) seriously discussed? It seems to me that the issues we have with PKI and certificate transparency could actually be mitigated very well if blockchain was seriously considered.
So I assume DLT is a generic/more general term for blockchain technology.
There's this thing called Zooko's triangle (https://en.wikipedia.org/wiki/Zooko%27s_triangle), basically the premise that an identifier can only have two of the following properties: secure, decentralized, human-readable.
But a blockchain can actually be used to square this triangle and get all three. The original example of this is the Namecoin project, a fork of Bitcoin using a lightly modified Bitcoin codebase which can be used as a key-value store.
The idea is that a blockchain can be used to create a decentralized database mapping keys (domain names) to values (which can be things like IP addresses, but also PKI trust anchors, similar to DANE). Thus this can eliminate the need for a CA. You can also use it to map human-readable names to .onion addresses. There is also root-of-trust transparency built in since the contents of the database is public and any changes are also public. The right to change a key-value entry belongs to the public key of the person who registered that key. Nobody can override this. Namecoin domain names use the .bit (unofficial) TLD.
It's a really neat technology and also makes things like censorship via domain suspensions, your registrar getting hacked, etc. infeasible.
Full disclosure, I previously worked on the Namecoin project, I authored the DNS resolution daemon and the technical specification for the format of DNS data in the KV store. Unfortunately public interest in the project has waned, and deployment is the real issue - you have to have the client software to be able to resolve these domain names. There is at least some conceivable possibility Tor might ship it in their bundle in the future though, so people could type example.bit in Tor Browser to get to a hidden service. The project remains active though.
Interesting and I can see some of the benefits, particularly in preventing DNS controllers from going to other CAs and making new keys, but... this all seems like a weird run-around for the MITM part.
Couldn't you exchange public keys with LetsEncrypt (in the web UI) and encrypt the response so you can't be MITMed? Why is http even an issue?
Because the attacker can simply request new keys. There’s nothing stopping them from going “hey LE, I need a new key! This is my domain, here is the challenge, give me my cert!” And LE will oblige, because as far as it can tell, they are you.
Edit: To be clear, this is a problem with a solution. But you asked why simply throwing a LE cert into the mix wouldn’t prevent the issue.
Ownership is already handled by checking DNS (and this thread covers a way to make that even more secure, which LE supports), and as far as I can tell neither has nothing to do with preventing MITM between LE and your servers.
And no, I don't mean throwing LE certs around to prevent MITM - this whole article is about the difficulties before having an LE cert, so that's necessarily excluded.
I'm wondering "why not client certificates". They're a well established way to stop MITM, seems like a simple choice for the ownership validation step.
Have you ever tried to explain key distribution/management to a normal person?
You should try it sometime.
Sure, but this is not particularly relevant when talking about a security product that you use to do key distribution and management (LetsEncrypt).
Explaining and guiding people through that is the whole point.
Could you please elaborate the reasons to implement this outside of DANE (RFC-7671) framework?
DANE assumes we can successfully deploy this to the entire Internet. It is unclear that's ever possible, and it's certainly not possible today. Lots of things would be great if you can deploy them, for example you wouldn't build QUIC on top of UDP since you can "just" deploy a new transport protocol - except nope, for the foreseeable future that's undeployable.
A public CA generally has a more sophisticated relationship with their network transit provider or (hopefully) providers and can get DNSSEC actually working as intended for them.
So this means mything.example's DNS services and some public CA both need working DNSSEC, but the visitors to mything.example, such as your mom's sister or some guy who just go into mything but isn't entirely clear whether Apple make Windows - do not need DNSSEc, for them everything works exactly as before, yet the certificate acquisition step is protected from third parties.
Would that help? It depends.
CAA is about preventing certificate mis-issuance, which is what happened in this attack. DNSSEC and CAA could have prevented this attack from being performed the way it was, by thwarting the MITM on ACME.
DANE is about changing the way certificates are authenticated. DANE makes it possible to authenticate certificates without getting them issued by a well-known CA. So CAA records are not particularly relevant to DANE. You can use DANE with certificates issued by a CA, which gives you two ways to authenticate the certificate; in this situation CAA secures one path and DANE the other.
I am one of the co-authors of the DANE SRV RFC https://www.rfc-editor.org/rfc/rfc7673 which is what XMPP would use. I don’t follow XMPP development so I don’t know if it has been deployed. I would like it if DANE were more widely used, but it’s not pertinent to this attack.
Why not? You could use "certificate usage" value 1 and (if the implementation does not neglect it) immediately notice that validation by CA disagrees with validation by DNS. That should be good enough, no?
Yeah. I used to be 100% in on DANE and against CAs. I'm still 100% for DANE but I now think DANE using existing CAs is the better option in many cases because it means things get CT logged. We don't have a DNSSEC transparency situation right now. OTOH there is one undersung issue with CAs, which is that Let's Encrypt isn't as universally available as people think (see the US embargo list) and that does potentially make access to the internet harder for some.
There are some use cases where DANE is actually winning real victories and is actually more viable than the existing CA infrastructure - site-to-site SMTP, for example.
Yeah, Viktor Dukhovni has been impressively energetic and persistent at improving the security of email.
I feel like packet size was and continues to be a major obstacle for DNSSEC. Do you know why the DNSSEC/DANE world hasn't simply acknowledged this and switched to requiring ECC?
It is trivial to fit several compressed curve points (i.e. signatures) in a single packet, whereas you can't even fit two RSA signatures in a minimum-safe-to-assume DNS UDP reply packet after accounting for padding and ASN.1 overhead.
I get the feeling that there is some faction that really hates UDP and they sort of hijacked the DNSSEC situation to use as a lever to force people to allow DNS-over-TCP.
That seems to be backfiring, however, and DNSSEC has wound up taking a bullet for the UDP-haters.
Many very-large networks simply can't afford for their DNS traffic to be exposed to TCP's intrusion-detection malleability and slowloris (resource exhaustion) attacks. These networks appear to be simply ignoring the "thou must TCP thine DNS" edict. DNSSEC is not a good enough carrot for them. I think ditching RSA would have been a more pragmatic choice than ditching UDP or skipping DNSSEC.
Dunno why there are so many foot-draggers failing to deploy better DNSSEC algorithms. I’m grumpy about SHA-1 in particular https://datatracker.ietf.org/doc/html/draft-fanf-dnsop-sha-l...
When I query vjhv.verisign.com I get a response containing four 2048 bit RSA-SHA-2 signatures in 1049 bytes which is well within the EDNS MTU for unfragmented UDP, so I’m not convinced the problem is as bad as you paint it. There have been problems with EDNS trying to use fragmented UDP, but that has been reduced a lot by newer software being more cautious about message size limits for DNS over UDP.
DNS needs TCP even in the absence of DNSSEC, because there are queries you cannot resolve without it. Some operators might convince themselves they can get away without it, but they will probably suffer subtle breakage.
> four 2048 bit RSA-SHA-2 signatures in 1049 bytes which is well within the EDNS MTU for unfragmented UDP
I was referring to the non-EDNS 512-byte limit.
Yes, you get ~2.5 times more with EDNS. Still, four records is not a lot.
> DNS needs TCP even in the absence of DNSSEC, because there are queries you cannot resolve without it.
Theoretically? Perhaps. Some would argue that connectionless DNS is valuable enough that people should not create those resource records. Before DNSSEC that was a working consensus. And with ECC it could be once again.
That's the opposite of the direction Internet cryptography is going, given hybrid PQC and classical systems.
The bloaty key/signature size is only a problem with the PQ encryption systems.
For signing only there are much more efficient PQ cryptosystems, with signatures around the same size as ECC. If DNSSEC ever adopts PQC it will be one of those systems.
Here are two of the earliest, and easiest to understand. There are much better ones now.
https://en.wikipedia.org/wiki/Lamport_signature#Short_keys_a...
https://en.wikipedia.org/wiki/Merkle_signature_scheme
Mostly, it's because very few serious engineering organizations deploy DNSSEC at all, so the best practices and tooling support aren't there.
Unfortunately the DANE SRV RFC is kind-of mismatched with how SRV and TLS work in practise. It requires the server to serve a certificate matching its own hostname (the hostname of the SRV target) rather than a certificate matching the expected host (the hostname that the SRV record was on). This is fine and secure if you use only DANE but if you want to use DANE with CA-issued certs it makes it somewhere between hard and impossible.
Note the owner of a SRV record is a service name not a host name.
There are a few reasons for this oddity: partly so it matches with DANE for MX records, partly to support large scale virtual hosting without reissuing certificates.
You should be able get a cert with subject names covering the server host name(s) and the service name(s).
Yep! This is one way of preventing this type of attack.