[tor-dev] Proposal 334: A flag to mark Relays as middle-only

[Roger Dingledine]

On Sun, Sep 12, 2021 at 12:17:37PM -0700, Neel Chauhan wrote:
>  If a relay has the MiddleOnly flag, we do not allow it to be used for the
>  following purposes:
>
>   * Entry Guard

While we're trying to be exhaustive here, "Directory Guard" might be a
good addition to this list. (But trying to be exhaustive is risky because
Tor's design will change over time and we'll forget to update this list.)

>   On an onion service host, when a INTRODUCE2 cell is received, if the
>   rendevous point has a MiddleOnly flag, the onion service host should close
>   the circuit and therefore not proceed with the protocol.

Two thoughts on this part:

(A) If we're teaching Tors to actively avoid touching these MiddleOnly
relays even when other people specify them, the rendezvous point
isn't the only one to look for. The next one that comes to mind is
the introduction point, i.e. if a client gets an onion descriptor that
lists an introduction point that has the flag, they would want to avoid
it. And now that we've got two examples, I bet there's a third, and even
if there isn't a third now, it's the sort of thing where future design
changes will forget to consider this part.

(B) There's a bigger problem here, stemming from desynchronized network
knowledge. For example, if my Tor doesn't think a relay has the MiddleOnly
flag, but your Tor thinks it does (e.g. because I have the consensus from
this hour and you have the one from last hour), then you'll refuse to
interact with me.

First, this situation can leak to me which consensus you're using,
which could build into other attacks. See this classic paper on this risk:
https://www.freehaven.net/anonbib/#danezis-pet2008

And second, this situation introduces hard-to-debug robustness issues,
which wouldn't be just a theoretical concern, since they would happen
each time the flag transitions on a given relay.

My suggestion would be to drop this idea of having Tors refuse to use
MiddleOnly relays in risky roles when other people specify them. We
already make sure to build our own path using relays we wanted to use,
before reaching those risky roles. Let's trust the other side to do it
too and not worry about it if it doesn't.

In the case of the two examples we've identified so far, the
attacker could use any relay they like in the next hop after that relay,
and we wouldn't know whether they're doing it. And for the rendezvous
point case in particular, it doesn't even need to be a relay that's in
the consensus right now (in part because we didn't want to get into the
information desync situation there too), so putting only this constraint
on what is an acceptable rendezvous point would be weird.

That is, I think these extra restrictions (avoiding the relays) would be
a slight improvement to security in theory, but I see that as outweighed
by the loss of robustness and by the other security angle (avoiding
letting people probe our internal network knowledge).

--Roger