This was a fix session of the quality programme. It took the email authentication checker, whose 2026-11-15 audit left eight defects open after two earlier fixes landed. All eight are fixed in one PR.
What was wrong
- Any non-empty
p=counted as a valid DKIM key. A 512-bit key, a truncated key,k=dsa,v=DKIM2andp=not*base64!were all reported as "valid public key". RFC 8301 s3.2 says verifiers must not accept signatures from RSA keys under 1024 bits, so a domain with one of these keys was told its DKIM worked when every signature fails. redirect=was counted even when the record also hasall. RFC 7208 s6.1 says the redirect is then ignored. Counting it could push a record past 10 lookups and produce a false PermError.- No test sat on any limit's edge. The only lookup-chain test used 11 lookups, so changing
> 10to>= 10, or moving the near-limit notice from 8 to 7, left every test green. - Smaller ones:
pct=50abcwas read as 50,adkim=xandsp=boguswere accepted without a word,ptrwas never flagged, a record ending inredirect=showed noallverdict at all, and the scope note saidmxhosts cost extra lookups. They don't: RFC 7208 s4.6.4 keeps those address queries out of the 10.
What changed
The checker now decodes p= and reads the RSA modulus length from the DER bytes. It accepts both the SubjectPublicKeyInfo form the RFC specifies and the bare PKCS#1 form some publishers use, and it checks that an ed25519 key is 32 bytes. Run live against hubspot.com, it reported the google selector's key as 1024 bits. openssl pkey on the same base64 also says Public-Key: (1024 bit). That key passes the floor and now gets a note to rotate to 2048. The same run shows hubspot.com's -all, read through its redirect= to _hspf.hubspot.com; before the fix the report showed no all verdict for that domain.
A new test file, rfc-boundaries.test.mjs, tests each threshold on its edge: 10 and 11 lookups, 7 and 8, pct 99 and 100, 512 and 1024 bits. Six seeded mutants, one per threshold or branch, now each fail a test. The page text no longer says that a record over the limit fails "for every message, for every receiver". Lookups are counted as the record is evaluated, so a message that matches an early term still passes.
What is still not at the bar
The checker still doesn't check the separate cap of 10 MX hosts per mx term, and it still guesses DKIM selectors from a fixed list of nine. The audit scores stay as they were. A later round re-audits the subject in a new file.