DKIM Replay Problem Statement

1.1. The problem

The presence of a DKIM signature serves as a basis for developing an assessment of mail received, over time, using that signature. That assessment constitutes a reputation, which then serves to guide future handling of mail arriving with a DKIM signature for that domain name. The presence of a validated DKIM signature was designed to ensure that the developed reputation is the result of activity only by the domain owner, and not by other, independent parties. That is, it defines a 'clean' channel of behavior by the domain owner, with no 'noise' introduced by other actors.¶

A receiving filtering system contains a rich array of rules and heuristics for assessing email, for protecting users against spam, phishing, and other abuses. DKIM therefore provides an identity that some systems can use for reputation assessment and prediction of future sender behavior.¶

During development of the DKIM specification, DKIM Replay was identified as only of hypothetical concern. However, that attack has become commonplace, particularly for systems:¶

• Attackers create, obtain access, or compromise an account at some Originator that signs messages with DKIM¶

• Attackers locate a Receiver that consumes DKIM to make a delivery decision. If the Receiver uses a reputation system with DKIM for delivery decisions, the attacker finds an Originator with a high reputation.¶

• They send an email from that account to an external account also under their control.¶

• This single message is delivered to the attacker's mailbox, giving them an email with a valid DKIM signature, for a domain with high reputation.¶

• They then post the message to a new and large set of additional recipients at the Receiver.¶

Internet Mail permits sending a message to addresses that are not listed in the content To:, Cc: or Bcc: header fields. Although DKIM covers portions of the message content, and can cover these header fields, it does not cover the envelope addresses, used by the email transport service, for determining handling behaviors. So this message can then be replayed to arbitrary thousands or millions of other recipients, none of whom were specified by the original author.¶

That is, DKIM Replay takes a message with a valid DKIM signature, and distributes it widely to many additional recipients, without breaking the signature.¶

• Further, a message used in a Replay Attack has the same attributes as some types of legitimate mail. That is, an individual, replayed message has no observable differences from a legitimate message.¶

Therefore, DKIM Replay is impossible to detect or prevent with current standards and practices. Simply put, email authentication does not distinguish benign re-posting flows from a DKIM Replay Attack.¶

ARC [RFC8617] is a protocol to securely propagate authentication results seen by Mediators that re-post a message, such as mailing lists. Because ARC is heavily based on DKIM it has the same "replay" issue as described in section 9.5.¶

1.2. Glossary

Modern email operation often involves many actors and many different actions. This section attempts to identify those relevant to Replay Attacks.¶

NOTE: This document is only Informative and omits the normative language defined in [RFC2119]. Mail architectural terminology that is used here is from [RFC5598] and [RFC5321].¶

[RFC5598] defines mail interactions conceptually from three perspectives of activities, divided into three types of roles:¶

Users:

This includes end-users, but also Mediators that re-post a message after delivery.¶

Services (Message Handling Service - MHS):

Moving a message from a single submission to its related delivery.¶

Administrative (ADministrative Management Domain - ADMD):

Covering independent operational scope, where functions of authorship, handling, and receiving can take place in any number of different ADMDs.¶

Also, as noted in [RFC5598], a given implementation might perform multiple roles.¶

It is useful to broadly identify participants in mail handling by functionality as defined in [RFC5598] as:¶

• Mail Submission Agent (MSA)¶
• Mail Transmission Agent (MTA)¶
• Mail Delivery Agent (MDA)¶

In addition, a user interacts with the handling service via a:¶

• Mail User Agent (MUA).¶

The following is a subset of the Mail Handling Services defined in [RFC5598] to be used in this document. The are summarized here for convenience:¶

Originator:

defined in Section 2.2.1. This is the first component of the MHS and works on behalf of the author to ensure the message is valid for transport; it then posts it to the relay (MTA) that provides SMTP store-and-forward transfer. The Originator can DKIM sign the message on behalf of the author, although it is also possible that the author's system, or even the first MTA, does DKIM signing.¶

Alias:

defined in Section 5.1. A type of Mediator user, operating in between a delivery and a following posting. The Alias replaces the original RCPT TO envelope recipient address but does not alter the content address field header fields. Often used for Auto-Forwarding.¶

ReSender:

as defined in Section 5.2, is a type of Mediator user, like an Alias; however the ReSender updates the recipient address, and "splices" the destination header field and possibly other address fields as well.¶

Mailing Lists:

defined in Section 5.3 is another Mediator; it receives a message and reposts it to the list's members; it might add list-specific header fields [RFC4021] e.g. List-XYZ: might modify other contents, such as revising the Subject: field, or adding content to the body.¶

Receiver:

defined in Section 2.2.4 is the last stop in the MHS, and works on behalf of the recipient to deliver the message to their inbox; it also might perform filtering.¶

Any of these actors, as well as those below, can add trace and operational header fields.¶

Modern email often includes additional services. Three that are relevant to DKIM Replay are:¶

Email Service Provider (ESP):

Often called a Bulk Sender - An originating third-party service, acting as an agent of the author and sending to a list of recipients. They may DKIM sign as themselves and/or sign with the author's domain name.¶

Outbound Filtering Service:

Rather than sending directly to recipients' servers, the Originator can route mail through a Filtering Service, to provide spam or data loss protection services. This service may modify the message and can have a different ADMD from the Originator.¶

Inbound Filtering Service:

The Receiver can also route mail through a Filtering Service, to provide spam, malware and other anti-abuse protection services. Typically, this is done by listing the service in an DNS MX record. This service may modify the message and have a different ADMD from the Receiver.¶

The above services can use email authentication as defined in the following specifications:¶

DomainKeys Identified Mail (DKIM):

Defined in [RFC6376], with a cryptographic signature that typically survives basic relaying but can be broken when processed by a Mediator. Further, DKIM Replay is defined in RFC6376 section 8.6.¶

Sender Policy Framework (SPF):

Defined in [RFC7208], is another form of message handling authentication that works in parallel to DKIM and might be relevant to the detection of a DKIM Replay Attack.¶

2.1. Basic types of flows

Direct delivery: In this case, email travels directly from the author's ADMD or the ADMD of their agent -- to the recipient's ADMD or their agent. That is, for origination and reception, any interesting creation or modification is done by agreement with either the author or the recipient. As such, these cases should have authentication that succeeds.¶

In this type of flow, SPF is expected to validate.¶

A DKIM Replay Attack uses a single message, sent through Direct delivery, and repurposes it.¶

Indirect Delivery:

This is mail involving a Mediator, producing a sequence of submission/delivery segments. While not required, the Mediator is typically viewed as being in an ADMD that is independent of the author's ADMD and independent of the recipient's ADMD.¶

2.2. Direct examples

ESP: An ESP is authorized to act on behalf of the author and will originate messages given a message body and a list of recipients, sending a different message to each recipient. Content address fields are typically restricted to just the address of that copy's recipient. The mail that is sent is typically 'direct', but the ESP cannot control whether an address refers to an alias or mailing list, or the like. So, the message might become indirect, before reaching the final recipient. : The bulk nature of ESP activity means that it can look the same as DKIM Replay traffic.¶

Outbound filtering: If the Author's domain has an SPF record that does not list this filtering service, SPF validation for the author's domain will fail. However, the ESP might produce an SPF record of their own and use their own SMTP MAIL FROM (return) address.¶

Inbound filtering:

Typically, an inbound filtering service will add the results of its analysis to the message. It might make other modifications to the message.¶

2.3. Indirect Examples

Indirect mail flows break SPF validation, unless the Mediator is listed in the SPF record. This is almost never the case.¶

Mailing List: The modifications done by a mailing list especially to the Subject: header field and the body of the message - nearly always break any existing DKIM signatures.¶

Alias (e.g., Auto-forwarder):

Typically, the envelope return (MAIL FROM) address is replaced, to be something related to the forwarder. A resender might add trace header fields, but typically does not modify the recipients or the message body.¶

3.1. Scenario

A spammer will find a mailbox provider with a high reputation and that signs their message with DKIM. The spammer sends a message with spam content from there to a mailbox the spammer controls. This received message is sometimes updated with additional header fields such as To: and Subject: that do not damage the existing DKIM signature, if those fields were not covered by the DKIM signature. The resulting message is then sent at scale to target recipients. Because the message signature is for a domain name with a high reputation, the message with spam content is more likely to get through to the inbox. This is an example of a spam classification false negative incorrectly assessing spam to not be spam.¶

When large amounts of such spam are sent to a single mailbox provider -- or through a filtering service with access to data across multiple mailbox providers -- the operator's filtering engine will eventually react by dropping the reputation of the original DKIM signer. Benign mail from the signer's domain then starts to go to the spam folder. For the benign mail, this is an example of a spam classification false positive.¶

In both cases, mail that is potentially wanted by the recipient becomes much harder to find, reducing its utility to the recipient (and the author.) In the first case, the wanted mail is mixed with potentially large quantities of spam. In the second case, the wanted mail is put in the spam folder.¶

3.2. Direct Flows

Legitimate mail might have a valid DKIM signature and no associated SPF record.¶

So might a Replay attack.¶

3.3. Indirect Flows

Example benign indirect flows are outbound and inbound gateway, mailing lists, and forwarders. This legitimate mail might have a valid DKIM signature, and SPF validation that is not aligned with the content From:¶

So might a Replay attack.¶

5.1. Include the SMTP RCPT-TO address in the DKIM signature

Since this information is different in the Replay, than it was in the original sending, locking it into the signature will make validation fail, if the value has been changed.¶

• This avoids Replay to destination addresses not anticipated by the DKIM signer.¶

• Indirect flows will fail, since forwarding involves rewriting the ENVELOPE-TO; however they already typically fail.¶

• This will detect DKIM Replays, but not distinguish them from all other forwarding.¶

• If a message has more than one addressee, should the signature cover all of them, or does this require sending one message per addressee? If it covers all of them, note that they might be on different systems, so that upon arrival, the RCPT-TO list will not include all of the original addresses¶

5.2. Count known DKIM signatures

This technique caches known DKIM signatures and counts them. Those above a certain threshold is considered DKIM replay.¶

• Since the same signature is being replayed many times, this might allow a receiving site with many mailboxes to detect whether a message is part of a DKIM Replay set, and to then suppress it.¶

• Mailing list traffic, aliases, and the like might also show up as duplicates. So this is only an heuristic, and might produce false positives.¶

• Caches have storage overhead, and uncertain TTLs.¶

5.3. Strip DKIM signatures on mailbox delivery

• Messages delivered to a mailbox lacking DKIM signatures can no longer be replayed.¶

• Has no effect when the receiving platform is collaborating with the bad actor, as the attacker would just avoid stripping the header fields.¶

5.4. Shorten DKIM signature key or signature lifetime

• If the key is no longer available through the DNS, the signature will no longer validate¶

• Alternatively express a validity period after which the signature is no longer valid¶

• Unfortunately, bad actors are quite good at taking action very quickly, and there is a limit to how much the window can be shortened, if the key is to have any utility for legitimate mail.¶

5.5. Add Per-hop signature, specifying the destination domain

Distinguish each forwarding hop by its own signature which permits each forwarding hop to specify the intended next destination ADMD. That intent can be verified to detect DKIM replay at the Receiver when the intended ADMD mismatches the current one.¶

• Messages with this kind of signature cannot be replayed down a different pathway, since the destination won't match.¶

• Requires every site along the path to support this spec, and to detect whether the next stop is making a commitment to follow the spec.¶

• If email goes to a site that does not support this behavior, traversing a naive forwarder remains indistinguishable from Replay.¶

• The time needed to change a global infrastructure such as email, to fully support a capability like this in every MTA is essentially infinite; therefore use of this approach must be narrowly tailored to scenarios that will adopt it and garner substantial benefit from it.¶