The M³WAAG DKIM Key Rotation Best Practices document (pdf) recommends a "sufficiently" random DKIM selector name so that it cannot be guessed by browsing the DNS. A literal quotation:
4.3 Key Selector Naming Scheme
Define a naming scheme for the DKIM key selectors that is both
meaningful for forensic analysis and is sufficiently random so the
keys cannot be easily guessed by browsing the DNS.NOTE: The selector naming scheme should also be designed to
mitigate the risk that attackers can easily predict the names of future
selectors and retrieve the associated keys. See Section 5 for a
description of the process for publishing keys for future use
This may be relevant for short 512-bit RSA keys, but it does not seem to make sense to me for longer, say 2048-bit, RSA keys. The DNS holds public keys which are not secret and can be discovered by reading just a single signed mail. Security by obscurity with very little security?
Why would a random DKIM selector name be better, when would it make sense to follow their recommendation?
Best Answer
I reviewed the document and found the author(s) don't understand DNS propogation of new entries. When updating old entries there are configurable cache times that can be several days. However, new entries need to be fetched from the authoritative name servers before the can be cached.
If keys are being rotated by the suggested process of rotating keys behind three CNAMEs, the there may be significant delays while cached entries are updated. This can be mitigated by dropping the TTL on record to be updated in the period before it is updated. The CNAME rotations may also be problematic in the case an emergency key rotation is required.
Randomizing the key names does provide some small measure of protection against the public key being retrieved in advance of use. Once the key is in use, I would assume that it could have been harvested for the purpose of generating an alternate signing key.