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
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.
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.
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.
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.
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.
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.
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...
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).
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
atlas would be good for detecting the time Pakistan announced 0/0 or whatever.
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.
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?
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.
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.
You should try it sometime.
Until the nation state tells a CA within their borders to issue a cert without publishing the CT logs, along with a gag order. If they don't comply they go to jail. You think Billy Bob Sysadmin wants to go to jail to protect a russian jabber server?
As the other commenter already pointed out, Firefox does not require this. Safari and Chrome do. This indeed is not directly applicable to this situation, since XMPP don't involve browsers. But for websites, the parents attack scenario is not applicable.
In the latter case, the malicious certificate + any CT log issued after that certificate should have been included are evidence of the attack that is easily verifiable and would likely cause browsers to drop the (edit) CT log unless provided with a very plausible excuse that isn't "we complied with a court/gag order".
Also, this would require the police etc. to compel 2-3 different entities: The CA and two log operators (I think one of them could be the CA itself).
Using different logs than normal would stand out, so it'd typically be two specific log operators that would have to be compelled to hide this. The US may get lucky and have jurisdiction over all of them, other countries are much less likely to.
Browsers would happily accept it and IIRC they don't check against the logs.
a) Browsers aren't the only clients that matter. Will any XMPP client even look at the CT log? Probably not.
b) Browsers are not going to drop all US-based certificate authorities. And if they tried they might just get their own national security letter (reminder: all major browsers are US-based).
You can't work around the government using centralized infrastructure.
With mild difficulty you can mirror the disk, net, and RAM of any "commodity" hardware.
Let's Encrypt security is a joke.
UPD: It looks like many other CAs also can be tricked to issue valid certificates to any attacker capable of doing MitM, so in general SSL infrastructure security is a joke which is proven by this incident.
Thanks to widely available HTTPS certificates, "evil hackers stealing your cookies on public Wi-Fi" is not a thing anymore.
We should definitely have a discussion about whether it's made active attacks more feasible, but I think the goal of making passive sniffing less trivial than it was before can be considered achieved.
Given that in the end, Let's Encrypt/ACME relies on PoP of a DNS domain, DNSSEC setup on the domain should have been a prerequisite.
In addition, an explicit opt-in should be required, for example, with a requirement for a CAA record pointing to LE.
---
As a side note, this whole situation leaves me with a really weird feeling.
On one end, I will always be grateful for LE enabling tls/https generalization.
On the other, I kind of feel betrayed and/or ashamed that LE/ACME almost willingly introduced protocol flaws weakening the whole CA infrastructure and that we (me included) didn't challenge it more when LE was introduced.
+1. crt.sh has RSS feeds for this.
> Limit validation methods and set exact account identifier which could issue new certificates with Certification Authority Authorization (CAA) Record Extensions for Account URI and Automatic Certificate Management Environment (ACME) Method Binding (RFC 8657) to prevent certificate issue for your domain using other certificate authorities, ACME accounts or validation methods
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?
> Monitor SSL/TLS certificate changes on all your services using external service
+1. High-risk targets should be aware of what certs are valid at any time, and be checking for those.
> Monitor MAC address of default gateway for changes
A more sophisticated attack could preserve the MAC address.
> "Channel binding" is a feature in XMPP which can detect a MiTM even if the interceptor present a valid certificate. Both the client and the server must support SCRAM PLUS authentication mechanisms for this to work. Unfortunately this was not active on jabber.ru at the time of the attack.
TIL.
I'd also add to these, consider an additional auth mechanism that doesn't rely on trusted CAs: Tor onion services, or something like Noise Pipes http://www.noiseprotocol.org/noise.html#noise-pipes