Re: [Mud] Intdir early review of draft-ietf-opsawg-mud-iot-dns-considerations-02

[David Lamparter <[e-mail]>]

On Fri, Dec 24, 2021 at 01:52:26PM -0500, Michael Richardson wrote:
> Normally, I like to reply with specific changes, but I haven't figured out
> what changes to make, and I also see that there are some XXX in my document
> that also reflect places I wanted to also make changes.
> So, please expect another reply with some specific text, once I've faulted
> enough pages back into the brain.

I'm noticing I communicated some of my thoughts quite poorly, probably
in part because I was somewhat unsure on distinguishing "comments" from
"review".  Apologies for that, I'll note to put a bit more time into
clearer communication on future reviews.

I also have really taken a wrong turn in basing some of my feedback on
thinking the names looked up by the device would match ACL entries in
the MUD file, particularly this first item:

>     > - the controller doing lookups, (potentially using the same caching resolver)
> 
>     > The MUD controller and the device using the same resolver is a neccssary
>     > requirement, but not a sufficent one.  Worse even, this is essentially a
>     > Time-Of-Check-Time-Of-Use security vulnerability.  An attacker may be able to
>     > distinguish device-originated DNS lookups from controller-originated lookups
>     > and give the controller a much longer list of replies while still allowing the
>     > device to function normally to cover their tracks.  It's fringe (if you can
>     > break DNS like that there's probably better things you can exploit), but still.
> 
> Yes, I agree that there is vulnerability here.
> 
> Having the MUD controller do DNSSEC validation helps a lot, if the
> manufacturer can be bothered to sign with DNSSEC.
> Or using DoH, DoT... as either way the attacker now has to compromise the
> authoritative servers.

In retrospect, I really regret bringing this perspective to the table.
It only makes sense from a "100% match/constraint" perspective.  But
with a make-before-break approach, the ACL being wider than the actual
usage by the MUD device is intentional.  Turning my paragraph on its
head, it might rather be prudent to repeat (for the benefit of network
operators) that a MUD file can be arbitrarily lax on its constraints.
Attacks on DNS can't widen it further than plain manufacturer
indifference to putting work into a good set of ACLs.


But after spending a bit more thought on this, I actually think I have a
good piece of proper "review" suggestion:  I think it would be immensely
useful to add/describe the perspective of what happens when addresses
associated with a service are 1. added, 2. removed, or 3. fully
replaced.  All the DNS caching and "perspective" effects come down to
the same differences, but they're corner cases to ops.  Moving services
around is core ops.

[-- continued below at anchor "A"]

> I feel that there is a lot more to say here.

And it really applies generally to using DNS names in ACLs.  I'm unaware
of any other document that went this direction, is there any?  RFC 8520
might've skipped over a seemingly innocuous but on closer look quite
nontrivial chunk of considerations here.

>     > I haven't really put extensive thought into this, but the only
>     > pedantically-correct solution to me seems to be to actually track the exact
>     > DNS responses that were given to the device.
> 
> Yes, that certainly can be done, and in the home router (which is also the
> DNS server for the device), then it occurs quite naturally.

[-- anchor "A", continuing on add/remove/replace]

Looking at these 3 temporal changes, the takeaway I've evolved to at
this point is that there are 2 overall patterns to get to "full
correctness":

- either the MUD controller shares the exact DNS->IP view that the MUD
  device has, ideally down to individual DNS lookup results.
  BUT: this implies the names listed in the MUD file must exactly match
  what the device looks up - and that's neither provided nor intended by
  8520.  Which really just renders this moot for further consideration.
  (It also requires controller and device to use the same DNS
  infrastructure.)

- or the entire thing instead moves to a make-before-break approach.
  The names in the MUD file receive address additions before the device
  sees them, and deletions after they expire out of device visibility.
  BUT: this implies the names listed cannot match what the device looks
  up/uses, and needs significant extra care by the device manufacturer.

The second item here I had kindof headed towards in my original review
when I mentioned using separate names, but I'll freely admit I hadn't
thought this to the end of being good ol' make-before-break.

Both approaches would solve the 3 "temporal operations", though in
almost antithetical ways:  construct a kind of atomicity to the changes
vs. split the operation into 2 or 3 stages.


[-- anchor "B", the following 3 paragraphs are redunant, left them here
just for clarity/explicitness]

Ultimately, again, this is really an issue of diametrically opposed
function of the name lookups.  Not sure which terms to use - "necessary
vs.  sufficient", "additive vs. subtractive", "permissive vs.
restrictive" - but it boils down to:

The MUD device starts with an empty set of endpoints and uses DNS to put
a few addresses into the set so it can pick something to talk to.  If
the set of addresses is *smaller* than the set of addresses that provide
the necessary service/function for the device to operate, that's
perfectly fine.

The MUD controller starts with allowing the unrestricted set of the
entire internet (equivalent to no MUD file too) and is using DNS to
narrow down this set.  If the set of addresses is *larger* than the set
id above, that's perfectly fine.

[-- end redundant]

> Note that the goal of MUD is not perfect mapping, but rather a reduction in
> risk.   As such I want to be very cautious about throwing away incremental
> improvements because they aren't perfect.

Fully agree.  Lacking sufficient context and knowledge, I unfortunately
have no clue what would lead to the best actual realized improvements.

>     > re 4.1. Use of IP address literals in-protocol
> 
>     > Nothing says the MUD file and device spec need to specify the target in the
>     > same way.  It might actually be a good idea to give the device a DNS name to
>     > resolve, but just list all the IP addresses in the MUD file.  The vendor should
>     > absolutely be able to do that in some cases at least (i.e. no highly dynamic
>     > changes to the addresses.)  Considering all else, this might even be one of the
>     > better approaches?
> 
> This is already possible.
> I think that it will lead to too many operational problems.
> Do you think that the document should deal with this in a pro/con section?

This paragraph of mine really came about from me thinking about the
make-before-break consideration above, though I hadn't arrived at that
myself just yet.  I suggested listing all the IP addresses only because
I was stuck with this "additive" use of DNS and took listing addresses
as "subtractive" converse.  The real point is to distinguish that
converse approach, not the distinction between use of DNS and addresses.

>     > On a sidenote, this also means it's wrong for the firewall to try to validate
>     > SNI names when the MUD file only lists IP addresses.
> 
> Yes, that's true.
> As we go towards ESNI, I hope that firewalls won't try to do this, as I think
> that *validating* SNI is a mistake.  Observing and auditing and logging if
> there are changes is a different situation though.
> I think it is reasonable for a MUD file to tell the MUD controller if the SNI
> should be visible or not (see draft-ietf-opsawg-mud-tls...)

Let me reword the review item:  for section 1, SNI considerations, I
think you should make clear the MUD file is making no commitment at all
that the names listed in the ACLs would match what the device ends up
negotiating a TLS connection to.  The DNS names in a MUD file are just
indirections to listing IP addresses.   So looking at SNI names is wrong
even when it's possible.

>     > Technically the MUD file could also list a *different* DNS name that is
>     > guaranteed to return a superset of addresses that might be used by the device.
>     > Whether this is a good idea I don't know.  (Also note this conflicts with
>     > monitoring a device's DNS lookups to track the actual responses it received.)
> 
> This definitely could be done.
> www-all.example.com, when www.example.com will return a geo-fenced address.

Exactly this.

>     > Ultimately though what I'd like to note here is that there's 2 different
>     > objectives here.  The device itself wants a useful subset of addresses to
>     > contact.  But the firewall needs the superset of anything the device might
>     > need to talk to.
> 
> Agreed.  I think that this paragraph is worth integrating into the document somewhere.

[this is what the earlier 3 paragraphs at anchor "B" are redundant with.]

[...some paragraphs cut...]

>     > 6.5. seems a no-brainer, maybe even too weak of a conclusion.  Devices /really/
>     > should use the DNS they get told by the network.
> 
> Yeah.  Are you asking for strong language here?

Yes, though to clarify, I should've inserted a "first" there.  Nothing
wrong with a fallback, if it's implemented correctly.  (And if a
particular network requires DNS lookups from devices to pass through its
DNS service to enforce some security purpose, it needs to be blocking
other DNS anyway.)

>     > 6.1. is the MUD file a protocol too?  It's not clear from the text.
> 
>     > (I do also disagree on the conclusion, in my naive world the vendor should
>     > really be able to list all possible IP addresses for the services, and doing
>     > so is likely the most robust way.  But I'm a "there is no cloud, just other
>     > people's computers" person, so maybe I expect too much of vendors.)
> 
> 6.1 is not about MUD exactly, but rather other protocols.
> Section 4.1 talks about this problem.   In my experience trying to fashion
> MUD files for things like printers, it's endemnic in the update protocol.

Ok - this absolutely needs wording clarification, because your current
wording can be read as recommendations to either "always use DNS names
on the device and MUD file", but also "always use DNS names, but this is
no statement about the MUD file".

>     > re 7. Privacy and 8. Security:
> 
>     >> Posession of intimate devices may cause embarassment.  Posession of personal
>     >> healthcare devices may open people up to attacks on their life.  Imagine some
>     >> dissident in Atlantis being assassinated on Princess Arielle's orders by way
>     >> of targeting something on the dissident's dialysis machine, pacemaker, or
>     >> whatever else.
> 
> May I use your Atlantis/Little Mermaid example?   :-)

Sure :D

> I am of the wrong age to really have experienced that film (too old for me,
> and it's too ancient to interest my kid... now teenager).  I don't remember
> there being dialysis machines in Atlantica.

Honestly I just didn't want to name any particular oppressive regime for
obvious political reasons.  I thought about fictional continents and
only associated to the movie from there...

>     > And lastly: "requirements for the devices to get access to network resources
>     > that may be critical to their continued safe operation." ... I guess this is
>     > more of a comment, but I certainly hope no hospital, metalworks, sewage works,
>     > or power plant ever depends on MUD functioning correctly to ensure *safe*
>     > operation.
> 
> Frankly, I hope that they do use MUD.

Oh they absolutely should.  I was trying to say they can't rely on it
for functional safety.  In most fields, safety regulations and
certifications will already require safe operation to be a property
inherent to the device as much as possible.  A MUD file is an extra
defensive layer, but if it fails open, that's no excuse for a device to
die in a "bad" way (e.g. because background internet traffic like SSH
portscans tripped/deadlocked something in its embedded firmware, and the
deadlock causes e.g. raw sewage to get into the drinking water supply.)


Cheers and Happy Holidays,

-David