Security audit 2026-10-17, finding W3 (low), stage 1. Resync read the
bindings and wrote the map with no lock (the selector table has tableMu);
two application changes at once could write in the order opposite to their
reads and leave the map stale - an application whose binding was missing
refused with reject_sender_login_mismatch until the next change. Now a
mutex on the Service holds the read, the write and the reload together.
Test: TestResyncRunsOneAtATime (eight at once, never two inside the
rebuild).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding A5/W2 (low), stage 1; the mockup is
7effc98. dmarc_reports belong to a domain by name, outlive the domain (no
FK), and a domain deleted and added again handed its old reports to
whoever it was assigned to then - the logs were moved to ids in 2.5.0 for
exactly this case, reports were not.
The cheap form of the plan: store.DeleteDomain deletes the domain's
reports (their records by the table's cascade) in the same transaction as
the domain. The confirmation page says so when the domain has reports
("delete the 12 DMARC reports received for it;"), counted by the new
store.CountDMARCReports; NewOutDomainDelete takes the count.
Tests: TestDeletingADomainDeletesItsReports (the domain added again has
none, another domain keeps its report and record, a second delete is
ErrDomainNotFound); TestOutDomainDeleteNamesWhatGoesWithIt with 12, 1 and
0 reports.
Docs: security.md § Roles; CHANGELOG; plan.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding W1, stage 1. HandleDomainDMARC took any
address validate.Email takes; AllowedRecipients opened every stored address
on the hostname, not only the domain's hosted one; checkReportDomain
cross-checked the claimed domain only against a dmarc-reports+<tag> local
part. So a domain user of tenant A could store ops@<hostname> and from then
on anyone could mail a report claiming tenant B's domain to it, stored and
shown under B and counted toward B's 500; or store B's hosted address and
open B's mailbox without B choosing hosted. TestAllowedRecipients asserted
ops@<host> was allowed - a choice, now reversed.
Now: a custom address on the hostname is refused on the form ("choose
SelfPost hosted"); AllowedRecipients adds a domain's address only when it
is that domain's own hosted address; checkReportDomain requires a tag and
the claimed domain to equal it. The live registry (checked 2026-10-17)
holds two hosted addresses and no custom one on the hostname, so nothing
there changes.
Tests: TestCustomReportAddressOnTheHostnameIsRefused (typed, another
domain's hosted, the domain's own hosted typed as custom; an address
elsewhere still saves); TestAllowedRecipientsFollowTheDomainAddresses
inverted plus another domain's hosted address stored; untagged addresses
rejected in TestIngestMessageRejectsTagMismatch; the ingest tests deliver
to tagged addresses.
Docs: security.md § Accepted risks (DMARC), guide § DMARC reports;
CHANGELOG; plan.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding M3, stage 1. inboundSession.RcptTo
appended every RCPT (the outbound session de-duplicates), and the end of
the message inserted a row per recipient one statement at a time inside
Postfix's callback; the cut to 10 000 rows every 500 counted only the log
tailer's rows. Postfix accepts a repeated recipient up to
smtpd_recipient_limit (1 000) and sends each to the milter, so one message
was a thousand rows, uncut until the six-hourly prune.
Now: the inbound milter folds a repeated recipient (case-insensitive, as
cleanup's been_here() with BH_FLAG_FOLD does, and as session.RcptTo does);
a message's rows go in one transaction (store.InsertInboundSpamMessage,
which replaces InsertInboundSpam in milter.InboundStore); the count and the
cut moved from the tailer into the store (countInboundWrites, an atomic
counter on Store) so the milter's rows, the tailer's and the tailer's
start-up re-reads all count. store.InboundSpamLogMaxRows is the one cap.
Tests: TestInboundRecipientRepeatedIsOneRow (1 000 RCPT in two cases, one
row); TestInboundJournalIsCutWhicheverWriterWrites (a small cap: a message's
rows and refusals together cross the mark, the table is cut); the tailer's
own counting test went with its counter. go test ./internal/store
./internal/milter ./internal/logtail (Windows-only Follow failures aside).
Also gofmt of a test line from the M2 commit.
Docs: security.md § Mail path, architecture.md; CHANGELOG; plan.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding M1, stage 1. overLimit took
domainOf(MAIL FROM) whatever the login, and MailFrom ran it with an empty
login: with smtpd_delay_reject at its default Postfix runs no sender
restriction at MAIL and sends the milter's MAIL event at once, so a
stranger on 465/587 with no AUTH, or another tenant's application, could
name a victim domain in MAIL FROM and take its slots for the 10-minute
in-flight TTL, get the victim's application answered 451, and have a limit
row (and the "a rate limit is refusing mail" notification) written under
the victim's domain.
Now overLimit returns at once without a login and takes the domain from
ApplicationByLogin (loginDomain, fail-open like the rest); recordRejected's
Domain is the login's too. The send log already counts a domain's messages
by the login's domain (store.CountMessages, loginDomainID), so the limit
and the count now agree. The null-sender special case is gone: every
message is held to its login's domain.
Tests: TestRateLimitIsNotReservedByAnotherSendersMailFrom (a stranger with
no login; another tenant's application) - fails on the old code at the
second MAIL FROM; the domain-limit tests sign in as app1 at example.com.
e2e domain_limit_not_reserved_by_a_stranger: a domain with a limit of one,
three unauthenticated MAIL FROM naming it on 465, then its own message
passes and its second is refused 451.
Docs: security.md § Mail path; CHANGELOG; plan.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding X1, stage 0: the supported-versions
table named 1.x while the product is at 2.14 - an operator would have
read 2.x as unsupported. Latest 2.x yes, earlier 2.x and 1.x no.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding S6 (low), stage 0 - proven from
outside: `ETRN <an inbound domain>` on port 25 answered "250 Queuing
started". smtpd_etrn_restrictions was empty (Postfix allows every client)
and fast_flush_domains = $relay_domains, so a stranger could flush the
deferred queue toward an inbound domain's upstream at will. Now
smtpd_etrn_restrictions=reject (etrn_cmd -> smtpd_check_etrn -> reject_code
554, Postfix 3.10.13 src/smtpd/smtpd.c) and smtpd_discard_ehlo_keywords=etrn.
e2e: testInboundRelay sends EHLO + ETRN for the inbound domain on port 25
and expects 554 and no ETRN in the EHLO reply (etrnPort25);
checkPostfixHardening reads both settings back. To be run on the server.
Docs: architecture.md § Postfix; CHANGELOG; plan.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding A4 (low), stage 0. The line logged when
a request carries two unprefixed session cookies printed r.URL.Path raw,
so GET /confirm/<token> wrote the token to the log (reachable only with
PANEL_COOKIE_SECURE=false, where a sibling host can plant the bare name).
loggedPath moves from internal/web to internal/web/auth as LoggedPath
(web imports auth, not the other way) and both lines - the origin check's
refusal and the duplicate-cookie one - go through it.
Tests: TestDuplicateCookieLineHidesALinksToken; TestLoggedPathHidesTokens
moved with the function. go test ./internal/web ./internal/web/auth.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding M5 (low), stage 0. senderSession.Header
appended every From:/Sender: value; Postfix bounds one header at 60 000
bytes but not their number, and the signer's 1 MiB cap runs after the
sender check, so an application could make it hold megabytes per
connection (the reader's test: a hundred lines, 6 MB). judge needs only
none / one / more, so two of each are kept; the refusal text joins what is
kept, as before.
Tests: "a hundred From: lines" in the judgement table (refused, more than
one From: line); TestSenderKeepsTwoFromAndSenderLines. go test
./internal/milter.
Docs: security.md (the sentence "a few headers each"); CHANGELOG; plan.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding M2, stage 0 - proven on the live server:
Postfix quotes a refused recipient in its reply in internal form (quotes
removed, "; < > =" kept: extract_addr, tok822_internalize), so
RCPT TO:<"; from=<ceo@…> to=<board@…> proto=ESMTP helo=<x>; "@domain>
spelled a whole tail inside the reply, and inboundRejectRe, whose reply
ended at the FIRST "; from=<", took the stranger's sender, recipient and
inbound domain for the row.
Now the reply is greedy and the tail anchored: "; from=<…> to=<…>
proto=\S+", a HELO name holding none of Postfix's NEUTER_CHARACTERS
(smtpd.c:1634, helo_cmd neuters ` <>()\";@`), and one-word SASL fields to
the end of the line (a `.*` SASL tail would have let a quoted MAIL FROM
spell the whole tail). Belt and braces: a reply that still holds
"; from=<" - a line with two tails - makes no row, so neither a recipient
nor an un-neutered HELO name can choose one. The quote/bracket/space
filter on from/to stays.
Tests: TestInboundRecipientCannotWriteTheRow (the live line's shape, and
with a forged MAIL FROM too); TestInboundHeloCannotWriteTheRow now feeds
HELO names as Postfix logs them (neutered) and expects no row for the
un-neutered ones. go test ./internal/logtail on Windows (the three known
Windows-only Follow failures aside).
Docs: security.md § Mail path; CHANGELOG; plan line ticked.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding I2 (= M4), stage 0. drop_links listed
the outbound maps, sasldb2, the key tree and mail.log but not the three
DMARC maps that postfix-config.sh creates as root with `[ -e ] || : >`
when DMARC_REPORTS_ENABLE=true; -e is false for a dangling link and the
redirect follows it, and /data/postfix is the panel's. Now the three are in
drop_links, and both map loops refuse to create through a link
(`[ -e ] || [ -L ] || : >`).
e2e: TestStartsOnA28Data plants a dangling /data/postfix/dmarc_transport
with DMARC reports on and expects the "removing the symbolic link" line,
no file at the link's target, and a regular map. To be run on the server
with the stage.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Security audit 2026-10-17, finding A2 (high), stage 0; the mockup is
2c27275. A stolen global session could set its own password on the user
form - no current password asked, the session kept - and then pass the
backup's reauth with it, or add a second global user with a known
password. Now the user form calls reauth (the backup's check: bcrypt
against the signed-in user's own hash, under auth.AllowPasswordChange) when
it adds a user and when a save sets a password, changes the e-mail (reset
links go there) or changes the role, the own account included; deleting a
user asks too, refused on its own confirmation page (renderUserDelete, a
Flash on UserDelete). Name and reach alone are not asked. Missing or wrong:
401 on the same page, "Your account password is incorrect.", nothing saved.
Tests: TestUserFormAsksForTheAccountPassword (create, each of password,
e-mail, role, own password, reach alone, delete); the user-form tests send
the field (withAccountPassword gives root a real hash and h an auth
module). go vet, go test ./internal/web/... on Windows.
Docs: security.md § Backup and domain export, guide § Users, help word for
word with the mockup; CHANGELOG [Unreleased] Security; plan line ticked.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Full security audit at 44bf09d (2.14.3): four zone readers, machine checks,
the live installation read-only, probes from outside. No critical; one high
(A2: a stolen global session sets its own password on the user form and so
passes backup-reauth), seven mediums (M1, M2, M3, W1, I1, I2, A1), the rest
low and nits; "checked and sound" per zone; doc drift; a seven-stage plan
with hours and a recommendation. The host's own findings are in dev/, not
published.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
mixandClaude Opus 5.5