False alarms are lethal for something as important as this.
Sorry...
False alarms are lethal for something as important as this.
Sorry...
That's not really true.
The canary is only "dead" and that channel is only useless until a new one is published. Unless the operator of the canary has decreed in advance that a particular time period elapsed should be considered "dead" regardless of future updates, then the canary is alive and well as soon as a new update comes out.
Here's the thing - a warrant canary is, by definition, a manual process. It should never be automated. A human has to resign and republish (or they are doing it wrong). And this means that sometimes they get published late.
In the 9 years that we (rsync.net) have been publishing our canary, we've probably missed our normal Monday morning publish 10 or 15 times. But as soon as we publish, with the affirmative canary and it's signature and language, etc., it's alive and well again.
Obviously this changes if someone is months late with an update.
That's what I was asking before - what is a reasonable timeslot for being "late"? Is it "months" or "weeks" or the <deltaManualUpdateT+1d> of canary x? In this case, SC was almost 3 times their weekly update rate late (December 5th to December 25th).
Once it gets to weeks, I think something else is going on.
By how much did you miss your Monday morning deadlines? Did you ever go a whole week without an update? Is someone tasked with doing it and someone else tasked with checking that it has been done (which is the way this should be set up, preferably with more than one person for the second role backed up with an automated process that sends out messages of impending aviary demise).
That signal has now gone off and I have no clue how you could put the genie back into the box without a time machine.
Then again, I don't know if the US can make you lie and coerce you to sign the canary. If so, then what you say makes sense.
"If the canary has not been updated in the time period specified by the host, users are to assume that the host has been served with such a subpoena."
That seems pretty clear to me.
Then, it stands to reason that if the canary gets updated again, you trust that everything is fine again, and that the owner didn't get served with an NSL. Therefore, it follows that if you trust canaries at all, you cannot stop trusting them if they fail to be updated for a time, except for that specific duration.
Please correct me if my reasoning is wrong anywhere.
Once dead they tend to stay dead. If you then prop up the canary afterwards and say it wasn't really dead, we just forgot to feed it for three weeks you're risking credibility at a minimum.
So either the operator is incompetent (which makes them a bad choice for trusting them with critical data) or the canary is now potentially under control of another party.
The problem I have with the whole setup is less one of whether or not this is a 'fixable' issue and we should redefine warrant canaries and give them a re-incarnation option, it's that if you think the chances of getting slapped with a NSL or warrant are large enough to have a canary system to begin with that you're dealing with customers that are - or should be - paranoid enough that any tampering with the system other than behavior as explicitly stated up-front should be un-acceptable.
Otherwise it becomes a question of where do we draw the line? A few weeks, a month or a year, what point is there to feed the canary at all if you can revive it at some arbitrary point in the future. After all the canary is a channel with very specific conditions attached to it, it now seems as though you are trying to redefine those conditions after the fact and that is the one property a thing like a warrant canary does not have.
If silent circle would have stated up front we will feed the canary weekly + a potential month in which we will take a holiday and everything is fine, otherwise we have been served with a warrant then you could call this operating within allowed parameters.
But I don't think you should be able to tack on such a clause afterwards.
The intra-period but within defined parameters risk is a different one from the post canary has given its signal one.
The way I understand the whole system to work is that you set up a fragile, fusible link that will blow exactly once to give its signal, and any further signals are suspect. Otherwise why agree on a protocol to begin with if you can then break that protocol at will?
I can see a case for changing the protocol assuming the previous one was not broken but once you violate the protocol I can't see a way to un-do that without making the whole concept meaningless as a result.
I don't think that conclusion is warranted. If the canary gets updated it could be any one of a number of things, one of which is that things are fine. But saying that the canary being updated is the same as you being compelled to lie misses out on a few critical options:
- presumably you set the time interval between updates short enough that your key or passphrase could not be compromised in that timeframe, so any interval longer may indicate a compromised private key
- it's only fine if the canary specs as defined up-front have not been violated (which in this case they have)
- canaries may be useless as a concept anyway because lawyers are human and hence fallible.
(That last one is a stretch, but with stuff like this it's either 1 or 0, not something in between and I don't see an iron clad guarantee there, merely an informed opinion.)