Since we're aiming for an Experimental document, it was discussed that it
makes sense to define that experiment. I'm proposing the following section,
and suggest it be the final section after Security Considerations but
before References.

For this section, I'm taking the stance "the best way to get the right
answer online is to give the wrong answer online." The one thing I know is
that what I've proposed below is wrong. It's just a matter of how, why, and
to what degree.

I'm also certain I'm missing critical design decisions, and the success
criteria probably needs some hard watermark (like 90% of mail that
fails/was subject to override now passes due to ARC alone).

Anyway, let's attack this straw man to figure out what the right form of
this section should be.

Seth

--------

16 Experimental Considerations

[[ NOTE TO WORKING GROUP: Should this section be for the IESG only to be
removed by the document editor, or should it stay with the document as long
as it’s experimental? ]]

It must be demonstrated that ARC actually solves the problem it is supposed
to - mainly, that ARC provides an effective signal to a Final Receiver that
allows messages indirectly delivered to properly be rescued after a DMARC
failure.

This section defines what success and failure look like, the protocol
design decisions that influence those criteria, and open issues that the
experiment should shine light on.

16.1 Design Decisions

16.1.1 Trace information is valuable

Because it is unclear exactly what information will be meaningful to Final
Receivers to make delivery decisions, it was decided that the protocol
would lean towards providing more information than less.

16.1.2 Chains can’t be restarted

Originally, it was expected that ARC Chains would be restartable. However,
this turned out to be a deeply complicated and implementation dependent
mechanism, for a use case that wasn’t clear would ever occur. The ARC-Seal
was necessary when restarting a chain to guarantee that the restarter knew
everyone who had been involved in the chain previously.

16.2 Success Criteria

Currently, many receivers have heuristic based overrides in place to
attempt to rescue mail from DMARC failures that they believe have been
delivered indirectly so as not to reject or junk mail their users
legitimately wish to receive.

ARC will be a success if, for ARC Sealed messages, receivers show equal or
better delivery rates than achieved through their overrides.

16.3 Failure Criteria

The intent of ARC is to be at most value-add and at worst benign. If ARC
opens up new vectors for abuse - beyond DKIM replay attacks and other known
issues (see:Security Considerations) inherent in the mail protocols ARC is
built upon - then ARC will be a failure.

16.4 Open Questions

The following open questions are academic and have no clear answer at the
time of the writing of this document. However, the ARC experiment should
gather the necessary data to conclusively answer some or all of them.

16.4.1 Value of ARC Seal

Because the ARC Seal was initially intended to provide context when
restarting a chain, now that restarting the chain is not something ARC
attempts to do, the value of the ARC Seal appears limited.

As part of the ARC experiment, data should be collected to show if the ARC
Seal provides value beyond the ARC Message Signature for either making
delivery decisions or catching malicious actors trying to craft or replay
malicious chains.

16.4.2 DNS Overhead

The longer an ARC Chain, the more queries that are needed to retrieve keys
to validate the Chain. It is not believed this will be a security issue
(see Security Considerations:DNS Attacks), but it is unclear how much
overhead will truly be added.

As part of the ARC experiment, data should be collected to better
understand the DNS impact of ARC Chains.

16.4.3 What trace information is valuable

There are several edge cases where the information in the AAR can make the
difference between message delivery or rejection. For example, if there is
a well known mailing list that ARC Seals but doesn’t do its own initial
DMARC checks, a Final Receiver with this knowledge could make a delivery
decision based upon the authentication information it sees in AAR[1].

Certain trace information in the AAR is useful in the construction of DMARC
reports.

Furthermore, certain large receivers believe the entire set of trace
information would be valuable to feed into machine learning system to shine
light on fraud or determine other signals related to message delivery.

However, it is unclear what trace information will be consistently valuable
for all receivers, regardless of size.

As part of the ARC experiment, data should be collected on what trace
information receivers are using that provide signal that affects
deliverability, and what portions of the trace data are left untouched or
provide no useful information.
_______________________________________________
dmarc mailing list
[email protected]
https://www.ietf.org/mailman/listinfo/dmarc

Reply via email to