Any casual reader of trivia will be aware that I care about my privacy and that I go to some lengths to maintain that privacy in the face of concerted attempts by ISPs, corporations, government agencies and others to subvert it. In particular I use personally managed OpenVPN servers at various locations to tunnel my network activity and thus mask it from my ISP’s surveillance. I also use Tor (over those same VPNs), I use XMPP (and latterly Signal) for my messaging and my mobile ‘phone is resolutely non-google because I use lineageos‘ version of android (though that still has holes – it is difficult to be completely free of google if you use an OS developed by them). Unfortunately my email is still largely unprotected, but I treat that medium as if it were correspondence by postcard and simply accept the risks inherent in its use. I like encryption, and I particularly like strong encryption which offers forward secrecy (as is provided by TLS, and unlike that offered by PGP) and will therefore use encryption wherever possible to protect my own and my family member’s usage of the ‘net.
Back in June of last year, I wrote about the problems caused by relying on third party DNS resolvers and how I decided to use my own unbound servers instead. I now run unbound on several of my VMs (in particular at my OpenVPN endpoints) and point my internal caching dnsmasq resolvers to those external recursive resolvers. This minimises my exposure to external DNS surveillance, but of course since my DNS requests themselves are in clear, any observer on the network path(s) between my internal networks and my external unbound servers (or between those servers and the root servers or other authoritative domain servers) would still be able to snoop. My DNS requests from my internal servers /should/ be protected by the OpenVPN tunnels to my unbound servers, but only if I can guarantee that those requests actually go to the servers I expect (and not to one of the others I have specified in my dnsmasq resolver lists). I attempt to mitigate this possibility on my internal network by using OpenVPN’s server configuration options
push “redirect-gateway def1”
push “dhcp-option DNS 126.96.36.199”
but my network architecture (see below) and my iptables rules on my external servers started to make this complicated and (potentially) unreliable. Simplicity is easier to maintain, and usually safer. Certainly I am less likely to make a configuration mistake if I keep things simple.
Part of my problem stems from the fact that I have two separate internal networks (well, three if you count my guest WiFi net) each with their own security policy stance and their own dnsmasq resolvers. Worse, the external network (deliberately and consciously) does not use my OpenVPN tunnels and my mobile devices could (and often do) connect to either of the two networks. Since both my dnsmasq resolvers point to the same list of external unbound servers this necessarily complicates the iptables rules on my servers which have to allow inbound connections from both my external ISP provided IP address and all of my OpenVPN endpoints. When thinking through my connectivity I found that I could not always guarantee the route my DNS requests would take, or which unbound server would respond. All this is made worse by my tendency to swap VPN endpoints or add new ones as the whim (and price of VMs) takes me.
Yes, a boy needs a hobby, but not an unnecessarily complicated one……
On reflection I concluded that my maintenance task would be eased if I could rely on just one or two external resolvers and find some other way to protect my DNS requests from snooping. The obvious solution here would be encryption of all DNS requests leaving my local networks (I have to trust those or I am completely lost). But how?
The first of these has never been an internet standard, indeed it has never been offered as a standard, but it was implemented, and publicly offered back in 2011 by OpenDNS. Personally I would shy away from using any non-standard protocol, particularly one which was not widely adopted, and very particularly one which was offered by OpenDNS. That company (alongside others such as Quad9 and Cleanbrowsing) market themselves as offering “filtered” DNS. I don’t like that.
Of the other two, both have RFCs describing them as internet standards, DoH in RFC 8484 and DoT in RFCs 7858 and 8310 but there is some (fairly widespread) disagreement over which protocol is “best”. This ZDNET article from October last year gives a good exposition of the arguments. That article comes down heavily against DoH and in favour of DoT. In particular it says that:
- DoH doesn’t actually prevent ISPs user tracking
- DoH creates havoc in the enterprise sector
- DoH weakens cyber-security
- DoH helps criminals
- DoH shouldn’t be recommended to dissidents
- DoH centralizes DNS traffic at a few DoH resolvers
The TL;DR is that most experts think DoH is not good, and people should be focusing their efforts on implementing better ways to encrypt DNS traffic — such as DNS-over-TLS — rather than DoH.
When people like Paul Vixie describe DoH in terms such as:
Rfc 8484 is a cluster duck for internet security. Sorry to rain on your parade. The inmates have taken over the asylum.
DoH is an over the top bypass of enterprise and other private networks. But DNS is part of the control plane, and network operators must be able to monitor and filter it. Use DoT, never DoH.
and Richard Bejtilch of TAO Security says:
DoH is an unfortunate answer to a complicated problem. I personally prefer DoT (DNS over TLS). Putting an OS-level function like name resolution in the hands of an application via DoH is a bad idea. See what @paulvixie has been writing for the most informed commentary.
I tend to take the view that perhaps DoH may not be the best way forward and I should look at DoT solutions.
Interestingly though, Theodore Ts’o replied to Paul Vixie in the second quote above that:
Unfortunately, these days more often than not I consider network operators to be a malicious man-in-the-middle actor instead of a service provider. These days I’m more often going to use IPSEC to insulate myself from the network operator, but diversity of defenses is good. :-)
and I have a lot of sympathy with that view. For example, that is exactly why I wrap my own network activity in as much protective encryption as I can. I don’t trust my local ISP.
In another discussion in a reddit thread, Bill Woodcock (Executive Director of Packet Clearing House and Chair of the Board of Quad 9) said:
[DNSCrypt] is not an IETF standard, so it’s not terribly widely implemented.
DNS-over-HTTPS is an ugly hack, to try to camouflage DNS queries as web queries.
and went on
DNS-over-TLS is an actual IETF standard, with a lot of interoperability work behind it. As a consequence of that, it’s the most widely supported in software, of the three options. DNS-over-TLS is the primary encryption method that Quad9 supports.
Whilst I might not like the way Quad9 handles its public DNS resolvers (and personally I wouldn’t use them), I can’t disagree with Woodcock’s conclusion.
My own view is that DoH looks very much like a bodge, and a possibly dangerous one at that. I’ve been a sysadmin in a corporate environment and I know that I would be very unhappy knowing that my users could bypass my local DNS resolvers at application level and mask their outgoing DNS requests as HTTPS web traffic. Indeed, when Mozilla announced its decision last year to include DoH within Firefox it caused some concern within both UK Central Government and the UK’s Internet Service Providers Association. Here I am slightly conflicted, however, because I can see exactly why that masking is attractive to the end user. For example, I sometimes run OpenVPN tunnels over TCP on port 443, rather than the default UDP to port 1194 for exactly the same reasons – camouflage and firewall bypass. And of course I use Tor. Mozilla reportedly bowed to UK pressure and did not (and have not) activated DoH by default in UK versions of Firefox. But it is not terribly difficult to activate should you so wish.
One of my main concerns over the use of a protocol which operates at the application layer, and ignores the network directives, is that those applications could, and probably would, come with a set of hard-coded DoH servers. Those servers could be hostile, or even if not directly hostile, they could be subverted by hostile entities for malicious purposes, or they could just fail. Mozilla itself hard-codes Cloudflare’s DoH servers into Firefox, but you could of course change that to any one or more of the servers on this list. The hard-coding of Cloudflare could cause Firefox users problems in future if that service were to fail. An article in the Register pointed to a failure in the F root server in February of this year caused by a faulty BGP advertisement connected with Cloudflare. As the Reg pointed out:
If a software bug in closed Cloudflare software can cause a root server to vanish an entire, significant piece of the internet then it is all too possible – in fact, likely – that at some point a similar issue will cause Firefox users to lose their secure DNS connections. And that could cause them to lose the internet altogether (it would still be there, but most users would have no idea what the cause was or how to get around it.)
Of course Mozilla isn’t the only browser provider to offer DoH. As the Register pointed out in November last year, both Google (with Chrome) and Microsoft (with Edge) are rolling out their own implementations. I’m with the Reg in being concerned about the centralisation of knowledge about DNS lookups this necessarily entails. If browsers (which after all are the critical application most used to access the web) all end up doing their DNS lookups by default to central servers controlled by a very few companies – and moreover companies which may have an inherent interest in monetising that information, then we will have lost a lot of the freedom, and privacy, that the proponents of the DoH protocol purport to support.
So DoT is the way to go for me. And I’ll cover how I did that in my next post.