# Changelog All notable changes to inbuxa-server. Versions follow the inbuxa suite: a series (`27.1`), its patches (`27.1.3`), and release candidates before a series ships (`27.1-rc.1`). The tag is the version with a `v`. Releases before 27.1 are dated (`2026.11.10`). The 2026 series is kept on the `release/2026` branch and gets security and data-loss fixes until 27.1 ships. Its releases are listed here too, by date. Each release's page has its container image and binaries: https://git.coffeylabs.org/inbuxa/inbuxa-server/releases. Security advisories: https://github.com/inbuxa/inbuxa-server/security/advisories. Every version PR adds that version's entry here. CI refuses a version this file has no entry for (`tools/release/version.py --changelog`). ## [27.1-rc.39] - 2026-11-21 27.1-rc.38 with one console addition. ### Added - The console's Settings › Network › Web links to Front ends, inbuxa Admin's page for `inbuxa:FrontEnds`: where inbuxa Admin and the webmail are served, the webmail's client secret, and other origins allowed to call the server from a browser. Until now only `inbuxa join` or the server's environment could set these. The page needs inbuxa Admin 27.1-rc.23; an older console hides the link. (#522) ### Upgrading - 27.1-rc.38 to 27.1-rc.39 rolls like any update, with no store layout step. ## [27.1-rc.38] - 2026-11-20 27.1-rc.37 with one fix. ### Fixed - Outbound DMARC and TLS reports wait up to 3 hours before their first delivery attempt, on purpose, so receivers don't get every report at the same moment. That wait was counted as delivery time, so the delivery-time graph and metrics showed hours for reports delivered on the first try. Delivery time now starts at the end of the planned wait. (#517) ### Upgrading - Reports already in the queue when you update still count their wait until they go out, at most 3 hours. - 27.1-rc.37 to 27.1-rc.38 rolls like any update, with no store layout step. ## [27.1-rc.37] - 2026-11-20 27.1-rc.36 with 2 security fixes from a private security review (1 high, 1 medium), two hardening changes and a fix for spam-training samples left behind. 27.1-rc.1 to 27.1-rc.36 are affected, and so is 2026.x up to and including 2026.11.20. The same security fixes are in 2026.11.20.1. ### Security High - WebDAV: an uploaded HTML or SVG file ran as a page of the server's origin when opened in a browser. Every WebDAV file answer now carries `Content-Security-Policy: default-src 'none'; frame-ancestors 'none'; sandbox` and `nosniff`; only raster images, audio, video and plain text are shown inline, everything else downloads. (GHSA-v4c6-686h-h927, #511) Medium - Multi-tenancy: a tenant administrator could make the server send requests to internal addresses through an ACME or DNS provider. A tenant's providers now reach only public addresses (HTTPS for ACME and HTTP-based DNS providers), RFC 2136 servers must be public, and fields built into URLs must be plain names. ACME errors carry only the problem type and detail. (GHSA-qmc9-xcrm-w836, #511) ### Changed - A tenant's OIDC directory reaches only public HTTPS addresses, discovered endpoints included. (#511) - Passkeys: a sign-in challenge or registration ceremony sent several times at once is used once. (#511) ### Fixed - Deleting a message whose spam-training sample was still held left the sample's record behind for good, also after the account was deleted. (#510) ### Upgrading - WebDAV: HTML, SVG, XML, PDF and untyped files now download instead of opening in the browser. - Tenant-owned ACME providers, DNS providers and OIDC directories at a private, loopback or plain-HTTP address stop working; move them to a public HTTPS address, or have a server administrator own them. - 27.1-rc.36 to 27.1-rc.37 rolls like any update, with no store layout step. ## [27.1-rc.36] - 2026-11-20 A tenant's disk usage now matches what its accounts store: calendar, contact and file writes are checked at the size they're charged, and destroying an account no longer upsets any tenant's count. ### Fixed - Destroying an account could reset the disk usage of an unrelated tenant that had the same number, so the tenant could store more than its quota until a recount. Destroying an account now takes its usage off its own tenant instead. (#507, fixes #499) - Copying or moving a single event or contact into another account didn't check that account's quota or its tenant's. (#507, fixes #499) - CalDAV, CardDAV and WebDAV writes were checked at the bare body length, not the size an item is charged, so each could go a few hundred bytes past a limit. Creating, updating, copying and moving items now check the charged size, and a refused file upload stores nothing. (#507, #508) - When a sharee, a delegate or an administrator wrote into someone else's account, the change in usage could be counted against the writer's tenant instead of the account's. (#507) - A tenant recount left out deleted members whose data is still kept. (#507) ### Upgrading - Run a tenant quota recount after upgrading if tenants' accounts have been destroyed, to correct counters the old destroy reset. - 27.1-rc.35 to 27.1-rc.36 rolls like any update, with no store layout step. ## [27.1-rc.35] - 2026-11-20 Tenant limits hold across many accounts and parallel writes, tenants verify the domain names they add, and a read-only config.json is reported clearly. ### Security High - Multi-tenancy: a tenant administrator could add any domain name, as a domain or as an alias, and the server then delivered mail that its other accounts sent to that name into the tenant. Names a tenant adds now have to be verified first, as described below. (GHSA-rvw7-gvh9-mx8g, #503) ### Changed - Multi-tenancy: a domain name a tenant administrator adds, as a new domain, a rename or an alias, has to be verified before the server uses it. The tenant publishes a TXT record at `_inbuxa-verification.` with `inbuxa-verification=` (the domain's `verificationToken`), then takes the name off the domain's `unverifiedNames`. Until then the name takes no mail and gives no addresses to send as. Names a server administrator adds need no record. Every name, aliases included, counts toward the tenant's `maxDomains`. (#503) - Items kept for undelete count toward their account's and tenant's disk usage until they're destroyed, expire or are restored; deleting still never fails. A held spam-training sample lets go of its message once the account's own copy is gone. (#502, fixes #500) ### Fixed - The accounts in a tenant could together be given more disk quota than the tenant has; four 5 GB accounts fitted in a 5 GB tenant. The total is now checked (accounts with no quota of their own share what's left), on restores and on domain moves into a tenant too. Thanks to BulTester. (#497, inbuxa-admin#257) - Parallel writes could each pass a tenant's count limits (accounts, domains and the rest) and together go over them. Count and disk limits are now checked under one lock per tenant. An object moved into a tenant by an update now counts against its limits, and DKIM keys a domain adds later count toward `maxDkimKeys`. (#501, fixes #498) - Encrypted mail was checked against quota before encryption, invitations without the notification they store, and a bigger vacation reply not at all. (#502, fixes #500) - Saving the data store settings when config.json is read-only (such as a Kubernetes mount) failed the whole request. That object is now refused with `forbidden`, saying the settings are managed outside the server and where. (#496) ### Upgrading - Domain names saved before this version are kept as they are, verified. If tenant administrators can manage domains on your server, review the domains and aliases your tenants have added. - Items deleted before this version aren't counted while they're kept; items deleted from now on are. - In a cluster, have tenants leave their domains alone until every node runs 27.1-rc.35: a domain saved through an older node during the roll loses its pending verification. - 27.1-rc.34 to 27.1-rc.35 rolls like any update, with no store layout step. ## [27.1-rc.34] - 2026-11-20 Tenants can no longer hand out more disk space than they have, listeners that can't open are retried and reported, backups straight to S3, metrics and alerts for Kubernetes, and signed images with SBOMs. ### Added - Release images are signed with cosign, keyless, on the registry and on the GHCR copy, with an SPDX SBOM per architecture and build provenance attested on GitHub. The SBOMs list the Rust crates in the binaries, not only the image's system packages. (#492, #493) - `--export s3://bucket/prefix` streams an export into S3 (or any S3-compatible store) without touching local disk, and `--import s3://bucket/prefix[/]` restores one, the newest finished export when no name is given. The bucket and keys come from `INBUXA_EXPORT_S3_*` variables. `--store-check quick|full` runs a store check to its end and exits with its result. (#491) - Metrics can be turned on from the environment: a password-protected `/metrics/prometheus` (`INBUXA_METRICS_PROMETHEUS_USERNAME` and `INBUXA_METRICS_PROMETHEUS_SECRET_FILE`), and OTLP from the standard `OTEL_` variables when the store sets none (`INBUXA_OTLP_HEADERS_FILE` reads the headers, such as a collector token, from a file). Three new gauges for alerts: `queue.oldest-message`, `cluster.lease-renewed` and `tls.earliest-expiry`. (#488) - A listener whose port is in use is retried for 30 seconds. Until it opens, `/healthz/ready` answers 503 naming it, and the console's cluster view lists it. `INBUXA_LISTENER_FAILURE=exit` stops the server instead when a listener still can't open. (#487, fixes #485) ### Fixed - A tenant administrator could give an account a bigger disk quota than the whole tenant has. The account's mail was refused as over quota while the webmail still showed most of its space free. Such a quota is now refused, whoever sets it, and JMAP and IMAP report the space the account can really use: its own quota or what's left in the tenant. Accounts already set too high keep their quota until it's edited. Thanks to BulTester. (#494, fixes inbuxa-admin#257) - An import into a fresh store refused to run, from a directory too, since 27.1-rc.16. (#491) - An import put blobs into the data store even when the imported settings name S3, so the restored server couldn't find them. (#491) - A live export could list blobs it didn't contain. (#491) - An S3 export or import could hang for good on a machine or pod limited to one CPU. (#491) - `/healthz/live` went through the anonymous request limit, so with the data store down the probe timed out and an orchestrator could restart every node at once. (#488) - Metrics: counters still at zero were left out of the scrape, gauges could wrap below zero, and in a cluster `queue.count` could read low. (#488) ### Upgrading - Nothing special: 27.1-rc.33 to 27.1-rc.34 rolls like any update, with no store layout step. ## [27.1-rc.33] - 2026-11-19 Passkeys that work on upgraded installs, a Passkeys link in the console's own Account section, a data-layer check, and certificates read from mounted directories. ### Added - `inbuxa --config --check` opens the data store the way a start does, writes, reads and deletes one small object in the blob store, and connects to the coordinator (NATS) once. It prints a line per part and exits, starting nothing. A part that can't be used is reported as `: : ...`, with the data-store exit codes from 27.1-rc.32. Before, a wrong S3 key or NATS password only showed after the start. (#473) - `INBUXA_TLS_CERTIFICATE_DIRS` names directories that each hold `tls.crt` and `tls.key`, as a Kubernetes TLS Secret mounts them. Their certificates are used for the names in them and aren't stored. Every node checks those files, and the files of file-backed stored certificates, once a minute, and reloads its certificates when they change, so a renewal no longer waits for a restart. (#477) - Console: the Account section gets a Passkeys link under Credentials, so people can manage their own passkeys without opening Directory › Accounts. Needs a console that has the page; an older one hides the link. Thanks to Stefan. (#481) ### Fixed - Passkeys: an install from before 27.1-rc.32 never got the passkey permissions added to its stored default roles, so nobody on it, not even the administrator, could list or add a passkey, and they couldn't be granted by hand either. Start-up now adds them once to the user, tenant administrator and administrator roles. A role that turns one off keeps it off. A test now fails if a later permission is missed the same way. Thanks to Faisal. (#483, fixes #478) - Two nodes starting together on a fresh store could both add the defaults, and the second came up without its HTTP listener. Adding the defaults now takes a short store lease, one node at a time. (#482, fixes #436) ### Upgrading - Nothing special: 27.1-rc.32 to 27.1-rc.33 rolls like any update, with no store layout step. - Passkeys start working on an upgraded install at the first start of 27.1-rc.33. Nothing needs changing by hand. ## [27.1-rc.32] - 2026-11-19 Passkey sign-in, a Logs page that can filter and follow, subjects in delivery history, a dark-mode logo, QR-code links in the spam filter, and the server-side pieces for running inbuxa on Kubernetes. ### Added - Passkeys: accounts can sign in with a passkey on the server's sign-in page, by autofill or a "Sign in with a passkey" button, in both the authorization-code and device flows. A passkey is enough on its own, with no TOTP code after it. Passkeys are added on a new page the server hosts, `/passkey/register`, after the password (and TOTP code) or an existing passkey, and managed over JMAP as `inbuxa:Passkey`. Removing one signs the account out everywhere. Passkeys are allowed by default; `passkeysAllowed` turns them off for the server, a tenant or a domain, and the session tells clients whether this account may add one. Accounts on an external directory (LDAP, SQL, OIDC) can't use passkeys, and IMAP, SMTP and the other mail protocols still use passwords or app passwords. Spec: `docs/spec/features/passkeys.md`. Thanks to Stefan. (#466, #468, #469, #470, #471, #474, fixes #334) - Delivery history: inbound and outbound lists show each message's subject. It comes from a debug-level event, `queue.message-subject`, capped at 255 characters, so it stays out of info-level log files and OpenTelemetry exports. Only full administrators can read it, and it is kept as long as history is. History stored before this release has no subject. Thanks to Faisal. (#452, fixes #228) - Logs: `x:Log/query` filters by level and above, component and time, and the Log tracer can write JSON lines (`format`: `text`, the default, or `json`). The Logs page reads text, JSON and a file holding both. Thanks to Stefan. (#455, #462, fixes #456) - Branding: a separate logo for dark mode, `logoDark` on domains and tenants and `logoUrlDark` server-wide. `/logo?theme=dark` serves it, or the ordinary logo where none is set; the built-in sign-in and RSVP pages ask for it when they show dark. Emails keep the light logo. Thanks to Faisal. (#459, fixes #410) - Spam filter: "Check links in QR codes" (`scanQrCodes`, off by default) reads QR codes in a message's images and sends the links through the same checks as its other links, adding the tag `QR_CODE_URL` (score 0.0 in the bundled rules, now 3.0.2-inbuxa.2). At most 4 images a message, 4 MB and 16 megapixels each, and 1 s in all. Thanks to Stefan. (#463, fixes #401) - OAuth token revocation (RFC 7009) at `POST /auth/revoke`, listed in both discovery documents. A revoked token is refused on every node until it would have expired. (#438) - `inbuxa --upgrade-store` upgrades the store layout and exits, and `inbuxa --upgrade-plan` says, read only, what it would do and whether a step needs every other node stopped. Both are for orchestrators that upgrade the store as a job of its own. (#448, #457) - `inbuxa --seed ` sets up a new store from a file of JMAP method calls, once, then exits. Spec: `docs/spec/features/store-seed.md`. (#451) - Secrets from files: `INBUXA_RECOVERY_ADMIN_FILE`, `INBUXA_WEBMAIL_CLIENT_SECRET_FILE` and, for the relay, `INBUXA_RELAY_JOIN_TOKEN_FILE`. Setting both forms, or naming a file that is missing or empty, stops the start. (#461) - `INBUXA_LISTENER_PORTS` moves or leaves out the default listeners a new store gets, for example `smtp=2525,submissions=4465,imaps=9993,pop3s=off,https=8443`. (#443) - `INBUXA_HTTP_USE_FORWARDED=true|false` overrides "Use X-Forwarded-For" from the environment. (#467) ### Changed - The container image runs with every capability dropped: the binary no longer carries `cap_net_bind_service` as a file capability. Docker is unaffected. Elsewhere, binding a port below 1024 needs `CAP_NET_BIND_SERVICE`, the `net.ipv4.ip_unprivileged_port_start` sysctl, or higher ports through `INBUXA_LISTENER_PORTS`, and a refused bind now says so. (#443) - A start that can't open the data store exits with a code per case: 78 bad configuration, 69 unreachable, 77 credentials refused, 76 TLS failed, 1 anything else, with "data store: " on its last line. Before, every case exited 1. (#472) ### Fixed - TLS: with no default certificate set, a connection with no or an unknown name got whichever certificate came first, and every handshake logged a warning. The server now uses the certificate for its own hostname (or a wildcard on its domain), else the first by name, and logs the warning once per certificate load. Thanks to khashmeshab. (#450, fixes #418) - Cluster: after a server was renamed, its old node stayed on the dashboard as "not responding" for a day. The old lease is now dropped once it has missed two renewals. (#460) ### Upgrading - Nothing special: 27.1-rc.31 to 27.1-rc.32 rolls like any update, with no store layout step. - Rolling back: an older server can't read delivery-history entries stored after this upgrade, drops a dark logo if it saves that branding record again, and ignores a Log tracer's `format` (it writes text). - Scripts that read the exit code of a failed start: see the data store codes above. ## [27.1-rc.31] - 2026-11-18 27.1-rc.30 with 3 security fixes from our own review of the suite (1 high, 1 medium, 1 low) and one hardening change. 27.1-rc.1 to 27.1-rc.30 are affected, and so is 2026.x up to and including 2026.11.18.2. The same fixes are in 2026.11.18.3. ### Security High - SMTP: one unauthenticated message could stall the server through DKIM2 verification, which took time quadratic in a signature's s= list. DKIM2 is now checked only when each signature has at most 4 s= values, the chain has at most 20 links and no DKIM2 field is over 64 KiB; otherwise the message has no DKIM2 result. (GHSA-8wjf-95c6-c4q3, #444) Medium - Mail parsing: header fields had no limit, so a large header block could exhaust memory. A part keeps at most 10,000 header fields and a message 262,144; over SMTP, more than 1,000 header fields in any part is refused with `552 5.3.4`. (GHSA-mvcp-vm37-f74j, #444) Low - Spam filter: an off-path host could forge Pyzor replies and mark mail as listed. Replies must now come from the Pyzor server queried and match the request, and request ids are random. (GHSA-4q7q-hg28-7666, #444) ### Changed - Spam filter: the URL homograph check looks up at most 10 distinct names per message, each once. ARC chains of more than 20 sets, or with a field over 64 KiB, are treated as broken. (#444) ### Upgrading - A message with more than 1,000 header fields in any part is refused over SMTP. ## [27.1-rc.30] - 2026-11-18 27.1-rc.29 with 4 security fixes from our own review of the suite (3 high, 1 medium). 27.1-rc.1 to 27.1-rc.29 are affected, and so is 2026.x up to and including 2026.11.18.1. The same fixes are in 2026.11.18.2. ### Security High - SMTP: one unauthenticated message could stall the server through DKIM verification, which took time quadratic in a signature's header list. At most the first 10 signatures are now checked, a signature listing more than 128 header fields is ignored, and a message with more than 1,000 header fields has none checked (dkim=none; refused with `550 5.7.20` under strict DKIM). (GHSA-v7ww-m523-wjfj, #439) - Spam filter: one unauthenticated message could exhaust the server's memory, since every punctuation mark became a token. At most 25,000 tokens per part and 100,000 per message are tokenized, and at most 1 MiB of HTML is parsed. (GHSA-gg2c-9hwc-59cm, #439) - Mail parsing: deeply nested message/rfc822 parts could overflow the stack and crash the server, again on every retry. Unencoded nested messages are parsed at most 20 levels deep; a deeper part is kept as an attachment. mail-parser 0.11.9 is now vendored for this. (GHSA-6245-jpv6-5pqm, #439) Medium - Mail store: a message with more than 65,535 MIME parts wrapped its stored part ids, and opening it exhausted memory. Such messages are now refused (`552 5.3.4` at DATA, NO on APPEND, an error from JMAP import), and ones already stored show only their well-formed parts. (GHSA-xm8r-p3w9-7cw8, #439) ### Upgrading - Spam padded past the token limits hides later text from the classifier and the URL checks; rules that read the text itself still see all of it. - A message with more than 65,535 MIME parts or nested messages is refused. ## [27.1-rc.29] - 2026-11-18 27.1-rc.28 with 6 security fixes from our own review of the suite (4 high, 2 medium), one of them also in inbuxa Relay. 27.1-rc.1 to 27.1-rc.28 are affected, and so is 2026.x up to and including 2026.11.18. The same fixes are in 2026.11.18.1. ### Security High - SMTP: one unauthenticated message could exhaust the server's memory through the DMARC check. A From header listing many addresses in different domains copied the whole message once per domain. Author domains are now deduplicated and checked without a copy; a From naming more than ten different domains fails DMARC with no lookups, and with strict DMARC verification is refused with `550 5.7.1`. (GHSA-jpf9-fq33-x6m6, #431) - IMAP and ManageSieve: empty quoted strings (`""`) didn't count toward the request size limit, so a client could grow memory without bound before signing in. They now count like any other argument. (GHSA-5f9w-xjcx-j9h2, #431) - Sieve: `replace` or `enclose` inside `foreverypart` crashed delivery on a message with nested parts, and enough such messages stopped local delivery for the whole node. `replace` now renumbers the remaining parts and the loop goes on over them; `enclose` ends the loop it runs in. (GHSA-jvv9-5xq7-64f5, #431) - HTTP: a client that stopped sending its body or reading its answer held a listener slot forever. Such requests and connections are now dropped after 60 seconds without progress. (GHSA-p33r-wr8j-hmw3, #431) Medium - SMTP: a signed-in user could inject SMTP commands and DSN header fields through an encoded line break in ENVID or ORCPT. ENVID, ORCPT and MAIL FROM `AUTH=` values that decode to anything but printable ASCII are now refused with `501 5.5.4`, and ENVID and ORCPT go out as xtext. inbuxa Relay does the same, and passes ORCPT on with its `rfc822;` type, so relayed recipients with ORCPT are no longer refused. (GHSA-c332-jfj8-p9w6, #431) - SMTP: on a listener offering AUTH without requiring it, AUTH after MAIL FROM made a transaction authenticated after its sender was checked anonymously. AUTH during a mail transaction is now refused with `503 5.5.1` (RFC 4954). (GHSA-7xrh-5v6g-hg45, #431) ### Upgrading - A client that sends AUTH after MAIL FROM must RSET first. - A From header naming more than ten different domains now fails DMARC. ## [27.1-rc.28] - 2026-11-18 27.1-rc.27 with 10 security fixes from our own review of the suite (6 medium, 4 low), three of them in inbuxa Relay, and two hardening changes. 27.1-rc.1 to 27.1-rc.27 are affected (inbuxa Relay from 27.1-rc.5), and so is 2026.x up to and including 2026.11.17. The same fixes, except the three for inbuxa Relay, are in 2026.11.18. ### Security Medium - SCIM: a deeply nested filter could stop the server. The filter parser had no nesting limit, so one request from a SCIM credential overflowed the stack and ended the server process. A filter longer than 8 KiB or nested more than 32 levels deep is now refused with `400 invalidFilter` (`invalidPath` for a PATCH path). (GHSA-6rvx-x4wr-558j, #419) - Account lock: a limited administrator could lock a fuller account and take over its mail by naming itself as delegate. Creating, changing and ending a lock now need every permission the account has, the rule that changing an account already follows. (GHSA-f3q6-j22f-7cf8, #419) - Mail flow: attachments in a few messages could stall the server. Each ZIP or Office attachment was unpacked up to 50 MiB, with no total per message, on the server's async workers. A message now gets one 10 MiB inspection budget across its bodies and attachments, unpacked data included, and at most 100 attachments are read; what's past either limit counts as "can't be inspected". Attachments are opened only when a rule checks content, the work runs off the async workers, and a failure while checking answers 451. (GHSA-qfcc-2hjm-7xjh, #419) - Relay: unauthenticated clients could hold port 25 sessions until inbuxa Relay ran out of file descriptors. Port 25 now takes at most `maxSmtpSessions` sessions at once (new setting, default 1000) and answers `421 4.3.2` past it, a client that doesn't read its replies for 5 minutes is disconnected, and the per-address limits count an IPv6 sender by its /64. (GHSA-2993-jvgq-9qvp, #419) - DLP: files inside Office packages, renamed text files and other archives went uninspected. Every file in an Office or OpenDocument package is now looked at, a file with no declared text type is read as text when it looks like text, and embedded objects, macro projects and non-ZIP archives (gzip, 7-Zip, RAR, tar and others) count as "can't be inspected". (GHSA-w2g2-f65h-wwhq, #419) - HTTP: the sign-in and consent pages could be framed by another site. Every HTML answer (sign-in, device sign-in, RSVP, and the console and account pages the server serves) now sends `X-Frame-Options: DENY`, `Content-Security-Policy: frame-ancestors 'none'` and `Referrer-Policy: strict-origin-when-cross-origin`. Every answer sends `X-Content-Type-Options: nosniff`, and every answer on a TLS listener `Strict-Transport-Security: max-age=15552000`. Headers set in `http.response_headers` take precedence. (GHSA-vxfg-3rh4-wc94, #419) Low - Branding: `/logo` was an open redirect to a domain's or tenant's URL logo. It now redirects only when a browser loads it as an image (`Sec-Fetch-Dest: image`) and answers 404 otherwise; the redirect isn't cached, and a logo stored as an `http:` URL is no longer used. (GHSA-49j4-h9f5-6h3x, #419) - Mail flow: an outside sender could keep a disclaimer, such as an external-sender warning, off their message by including its text. Incoming mail now always gets the disclaimer; outgoing mail still skips one that's already there. (GHSA-6p5j-4m95-c6qh, #419) - Relay: with the link down and 10,000 events waiting, each new connection rewrote and synced inbuxa Relay's whole event file. The oldest thousand are now dropped at once. (GHSA-j2qx-wxvv-5x6h, #419) - Relay: mail still being received wasn't counted toward `spoolMaxBytes`. It now is, and a message that would take the spool past it, or the spool's disk past 95%, is refused with `452 4.3.1` at the end of `DATA`. (GHSA-wcq3-6qc3-f5w2, #419) Hardening - DLP: detectors no longer crash on digits from other scripts (Arabic, Persian, Hindi, fullwidth). A message holding them ended the sender's session; it's now checked and sent normally. (#419) - Autoconfig and autodiscover: answers carry `Content-Security-Policy: default-src 'none'; sandbox` and `X-Content-Type-Options: nosniff`, so they can't run as a page on the server. (#419) Found by Coffey Labs. ### Changed - The container image bundles inbuxa-migrate 27.1-rc.16: `import jmap` no longer sends the inbuxa profile's API key to the source server, and certificate checks no longer trust the distrusted emSign intermediates. Migrations the server runs itself were not affected. (#422) ### Upgrading - After a browser has visited the server over HTTPS, it will also use HTTPS for `http://:8080`. Open the plain listener by IP address, or set your own `Strict-Transport-Security` in `http.response_headers` (e.g. `max-age=0`) if you rely on it by name. - To show the server's pages in a frame (a portal, say), set your own `Content-Security-Policy` with `frame-ancestors` in `http.response_headers`. - "Can't be inspected" now covers more attachments: gzip, 7-Zip, RAR, tar, xz, bzip2, zstd and CAB files, Office files with embedded objects or macros (`.docm`, `.xlsm`), and attachments past a message's 10 MiB inspection budget or its first 100. Rules on that condition act on them now. - A URL logo shows only where a browser loads `/logo` as an image, as the sign-in page and the console do. Opening the link directly gives 404, and browsers that don't send `Sec-Fetch-Dest` (Safari before 16.4) show the built-in logo. Logos stored as `http:` URLs are no longer used; set an `https:` URL or an image. - A role that can lock accounts can no longer lock, change the lock of, or unlock an account holding permissions it lacks, administrators included. - SCIM filters longer than 8 KiB or nested more than 32 levels deep are refused. Identity providers send far less. - inbuxa Relay: `maxSmtpSessions` (default 1000) caps port 25 sessions across all senders, and the per-address rate limits count an IPv6 sender by its /64, so a large sender using many addresses in one /64 shares one budget. The console doesn't show `maxSmtpSessions` yet; set it with `inbuxa:Relay/set` (`settings/maxSmtpSessions`). An older relay ignores it. ## [27.1-rc.27] - 2026-11-17 27.1-rc.26 with 2 security fixes from a private review of the server (1 medium, 1 low) and four hardening changes. All versions up to and including 2026.11.11, and 27.1-rc.1 to 27.1-rc.26, are affected. The same fixes are in 2026.11.17. ### Security Medium - HTTP: behind a reverse proxy, a client could choose its own address. With forwarding headers trusted, the server read RFC 7239 `Forwarded` before `X-Forwarded-For`, so a proxy that writes only `X-Forwarded-For` passed a client's own `Forwarded` header through. It now reads one header, `X-Forwarded-For` unless `INBUXA_HTTP_FORWARDED_HEADER=forwarded`, and a loopback address in it counts as this host only when every hop after it is loopback too. (GHSA-hxp3-8c95-4q5v, #413) Low - Multi-tenant: a tenant administrator could download any blob by its hash. Inside a tenant, `fetchAnyBlob` now reaches only blobs the tenant holds: linked by one of its accounts, or in a queued message the tenant sees. Tenant queue downloads keep working. (GHSA-x596-5p95-2rm3, #413) Hardening - TLS: rustls-webpki 0.103.17, webpki-roots 1.0.12 and webpki-root-certs 1.0.12 stop trusting the distrusted emSign intermediates (RUSTSEC-2026-0381, RUSTSEC-2026-0382, RUSTSEC-2026-0383). (#413) - Tenants: no tenant policy grants writing system Sieve scripts, which can run programs as the server's user. (#413) - OAuth: with `requireClientRegistration` off (it's on by default), a client nobody registered must use PKCE (S256) and may only redirect to a loopback address or a private-use scheme (RFC 8252). (#413) - Downloads: a requested type that isn't a valid header value falls back to `application/octet-stream`, and a response with a bad header value is a 500 instead of a dropped connection. (#413) These came from a private security review; thanks to its author. ### Upgrading - If your reverse proxy sends only RFC 7239 `Forwarded`, set `INBUXA_HTTP_FORWARDED_HEADER=forwarded`; otherwise client addresses read as the proxy's. nginx, Caddy and Traefik send `X-Forwarded-For` and need nothing. - With `requireClientRegistration` off, a web application signing in by OAuth has to be registered now. ## [27.1-rc.26] - 2026-11-17 The bundled inbuxa-migrate moves to 27.1-rc.15. ### Changed - The container image bundles inbuxa-migrate 27.1-rc.15: upstream v1.1.0's dependency updates, and `export` no longer sends a message again on every run when the target keys it differently, as with a message carrying two `Message-ID` headers. (#411) ## [27.1-rc.25] - 2026-11-16 A TLS report counts once on the dashboard, however many policies it covers, and Deliverability stops failing SPF for a container's own addresses. ### Fixed - TLS reports: a report logged one incoming-report event per policy, so a report covering both an MTA-STS and a DANE policy counted as two reports on the dashboard, and as two with warnings when both had failures. It now logs one event per report, with the totals summed over its policies and each policy's figures kept in the event. Thanks to khashmeshab. (#406, fixes #402) - Deliverability: with no source addresses set, the check tested what the server's name resolves to, and in a container /etc/hosts maps that name to the container's own addresses. On an installer's container install it failed SPF for 172.31.253.10 and fd69:6e62:7578::2, which no receiver ever sees. Private, shared, loopback, link-local and unique-local addresses are left out now, and when those are all the name gave, DNS itself is asked. Thanks to faisal. (fixes #408) ## [27.1-rc.24] - 2026-11-16 The sign-in page no longer stays behind in the browser history. ### Fixed - Sign-in page: after a successful sign-in the page sent the browser on to the app with `location.assign`, which left it in the history. In the console and webmail installed as apps, Back then landed on a spent sign-in form, and signing in there failed. It now replaces itself. Thanks to khashmeshab. (#404, fixes #403) ## [27.1-rc.23] - 2026-11-15 Delivery history keeps inbound and outbound mail apart, automatic DKIM makes a DKIM2 key as soon as it is ticked, and release downloads are signed. ### Added - Delivery history: inbound and outbound lists can be scoped by kind. `x:Trace/query` takes a `kind` filter (received, submitted, remote, local, any), worked out from each trace's events, so older history is covered too. Inbound defaults to mail received from other servers and Outbound to remote deliveries, so mail your users send over 587/465 and local deliveries no longer mix in; a Show filter brings them back. Thanks to Faisal. (#398, fixes #397) - A `spam.result` event (info level) records the spam verdict, score and the rules hit, highest points first, so they appear in an inbound trace. (#398) - Release downloads come with `SHA256SUMS.minisig`, a minisign signature of `SHA256SUMS`. (#396) ### Fixed - Automatic DKIM: ticking a new signing algorithm on a domain already set to Automatic, such as DKIM2 next to DKIM1, made no key until the next rotation. The save now makes it. Unticking one still takes effect at rotation. Thanks to Stefan. (#399) ## [27.1-rc.22] - 2026-11-13 Fixes from the community: Check now reads fresh DNS, Automatic DNS publishes newly enabled record types, and the dashboard counts DMARC warnings the way the reports page does. ### Changed - The container image bundles inbuxa-migrate 27.1-rc.14: upstream v1.0.16's Exchange EWS and Graph fixes, and the POP3, Fastmail, Zoho and Zimbra sources. (#393) ### Fixed - Deliverability: Check now used the same cached DNS answers as mail traffic, so a check right after fixing reverse DNS or adding a DMARC record could still show the old state. A manual check now skips the caches. The daily check and SMTP keep them. Thanks to Stefan. (#387, fixes #385) - Automatic DNS: adding a record type to `publishRecords` on a domain already set to Automatic queued no DNS task, so the new records were never published until the domain was switched to Manual and back. The save now publishes them. Thanks to Stefan. (#389, fixes #372) - DMARC reports: the dashboard marked a report "with warnings" when any row failed SPF or DKIM, so forwarded mail that passed DMARC on DKIM counted as a warning. A warning now means a row failed DMARC, as on the reports page. (#392, fixes #391) ## [27.1-rc.21] - 2026-11-11 DMARC and TLS reports reach the server on every domain without a postmaster account for each, and the sign-in page can match the app that opens it. ### Security One advisory (GHSA-c3c4-2c6m-ffrc), overall severity High: records under an active legal hold could be destroyed on schedule (spoliation). 27.1-rc.1 to 27.1-rc.20 are affected. - Legal hold: an account deleted while undelete still retained it now has its archived mail frozen by a covering hold, instead of being purged on schedule. (#377) - Legal hold: the audit and journal records of such a deleted-but-held account are kept under the hold, instead of being purged once past their retention floor. (#381) ### Added - The sign-in page takes `theme=light` or `theme=dark` from the app that opens it, so it matches the console or webmail instead of following the device. Without the parameter it follows the device as before. (#375, #376) ### Changed - With no default folders configured, a new account also gets an Archive folder. A folder list an admin set is left as it is, and existing accounts aren't changed. (#378, fixes #374) - Mail to a report address (`postmaster@*` by default) on one of the server's domains is taken even when no account has that address, as long as the sender hasn't signed in. A DMARC, TLS or feedback report in it is read as usual; anything else is refused at DATA with the same "Mailbox does not exist" as before. Extra domains no longer need a postmaster account for their reports. (#379, fixes #373) ## [27.1-rc.20] - 2026-11-11 The audit log, the journal, legal holds and undelete's notes move to tables of their own. ### Changed - Store layout version 3. The audit log, the journal, legal holds and undelete's notes move out of the settings table into tables of their own (column families on RocksDB): `ib_audit`, `ib_journal`, `ib_holds` and `ib_undelete`. The first node to start copies them, checks the copy row for row against the original, and only then removes the originals. A run that's cut short picks up where it stopped. See https://docs.inbuxa.org/install/updating/#store-layout (#368) - Exports and backups have a `compliance` family for the same data, on by default. Before this release it travelled with `registry`, so `EXPORT_TYPES=registry` now copies the settings without it. See https://docs.inbuxa.org/run/backup/ (#368) - Relayed inbound mail follows the settings in force now. A relay link judged its messages with the settings it had when it connected, so a change to a domain's directory, the RCPT and DATA rules, the spam filter or sender authentication didn't reach mail arriving through a relay until the link reconnected. (#366, #367) - A recipient refused because a domain's LDAP or SQL directory had no answer for it is logged at info level, as `smtp.directory-no-answer`, with the reason and the filter or query that ran. (#364, #365) ### Before you upgrade - This upgrade can't roll. Stop every node, start one on 27.1-rc.20 and wait for "Upgraded the store layout to version 3" in its log, then start the others. Mail is down for those few minutes. - Back up the data store first. An older release refuses the store afterwards (on RocksDB it can't open it at all), so going back means restoring that backup. - Run this with inbuxa-admin 27.1-rc.13 and inbuxa-webmail 27.1-rc.3. ## [27.1-rc.19] - 2026-11-10 Once a blob store change is settled, the server can clean up the old store. ### Added - Clean up the old store after a blob store change. Once a change is copied, the old store still holds a copy of every blob it had. A cleanup deletes them from it, each one only after reading back its copy in the new store and checking it: message content against its id, anything else byte for byte against the old copy. A blob that doesn't check out stays where it is and is listed. A preview (`dryRun`) counts and deletes nothing. The cleanup needs `confirm` set to the change's id, because backups of the data store taken before the change read their blobs from the old store and won't restore in full afterwards. It's refused while a node still has its blob store in the old place, when the two stores overlap, and while a copy runs. Over the API it's `"action": "cleanUp"` on `inbuxa:BlobStoreChange`; in inbuxa Admin 27.1-rc.12 it's **Clean up the old store**, and in inbuxa-cli 27.1-rc.2 it's `store clean-up`. (#342, #357, #358, #359) ### Before you upgrade - Changes settled from this release on keep the old store's settings so it can be cleaned up. Changes settled before it can't be cleaned up from the server; remove their old store by hand, as before. - No change to the store's layout; it stays at version 2. Upgrading from 27.1-rc.16 or older still runs the upgrade in 27.1-rc.17, so read its notes first. - Run with inbuxa-admin 27.1-rc.12 and inbuxa-webmail 27.1-rc.3. ## [27.1-rc.18] - 2026-11-10 A blob store change can be planned ahead and copied before the setting moves, and snapshots stay listed while a change isn't settled. ### Added - Plan a blob store change ahead. Give the new store's settings, and the server checks it can write there, then copies everything into it while the current store stays the blob store. Mail keeps going to the current store meanwhile, and a blob deleted there is deleted from the new store too. Once that copy ends, "apply" writes the planned settings to the blob store setting, credentials included. After the nodes restart on it, the last copy transfers only what arrived in between. One change can be planned at a time, and none while another isn't settled. Over the API it's `inbuxa:BlobStoreChange` with `target` on create and `"action": "apply"`; in inbuxa Admin 27.1-rc.11 it's **Plan a change**. (#341, #354, #355) - Settings › Storage links to inbuxa Admin's pages for the store check, snapshots and blob store changes. (#350) ### Fixed - While a blob store change wasn't settled, the snapshots already on disk went missing from the server's view: none were listed, deleting one answered `notFound`, and start-up didn't clean up a partial one. They're listed and can be deleted again; only a new snapshot is refused. (#351) ### Before you upgrade - No change to the store's layout; it stays at version 2. Upgrading from 27.1-rc.16 or older still runs the upgrade in 27.1-rc.17, so read its notes first. - Run with inbuxa-admin 27.1-rc.11 and inbuxa-webmail 27.1-rc.3. ## [2026.11.10] - 2026-11-10 A bug fix for the 2026 series. No security fixes beyond 2026.11.9. ### Fixed - Store snapshots: while a blob store change wasn't settled, the snapshots already on disk went missing from the server's view. `inbuxa:StoreSnapshot/get` listed none, deleting one answered `notFound`, and partial snapshots weren't cleaned up at start-up. A new snapshot is still refused until the change is settled, but the existing ones are listed, can be deleted, and are cleaned up. Nothing on disk was lost. (#352) ### Upgrading - Nothing to do. Rolling back to 2026.11.9.2 is safe. ## [2026.11.9.2] - 2026-11-09 A bug fix for the 2026 series. No security fixes beyond 2026.11.9. ### Fixed - CalDAV/CardDAV: copying or moving a whole calendar or address book into another account skipped every entry whose collection id in the old account matched the destination collection's id in the new one. A copy came out empty, and a move deleted the source while those entries stayed behind, listed nowhere. A copy onto a calendar or address book of the same name in another account was also refused with 502. (#347) ### Upgrading - Nothing to do. Rolling back to 2026.11.9.1 is safe. ## [27.1-rc.17] - 2026-11-09 The first change to the store's layout, a fix for calendars and address books that lost entries when copied between accounts, and the bundled inbuxa-migrate 27.1-rc.12. **This release upgrades the store when it first starts, and that can't be undone.** Once it has run, 27.1-rc.16 and older refuse to start on the store. The only way back is a backup taken right before the upgrade. ### Added - Store layout upgrades. A release can now change how the store is laid out, in steps, at startup. The log says when a step starts and ends and what it changed, and warns before it starts that it can't be rolled back without a backup. Every start logs the store's layout version. In a cluster, one node runs the step and the others wait for it. A running node that finds the store upgraded by a newer release stops, so it can't keep writing the old layout. (#345) - Layout version 2. On PostgreSQL, MySQL and SQLite, a rate-limit counter keeps its expiry in its own row instead of a second row beside it, so a counted request writes one row, not two, and expired counters are cleared with one statement. The counter table gets a new column at startup and the upgrade moves existing counters over. On RocksDB and FoundationDB only the version changes. (#345) ### Changed - The container image bundles inbuxa-migrate 27.1-rc.12. (#339) ### Fixed - CalDAV/CardDAV: copying or moving a whole calendar or address book into another account lost entries, as described under 2026.11.9.2. (#346) - Rate-limit counters write their expiry marker once per time bucket, not on every request, on every backend. (#344) ### Before you upgrade - Take a backup right before you start this release. A nightly backup isn't enough: restoring it loses the mail that arrived since. For PostgreSQL or MySQL, dump the database. For RocksDB or SQLite, copy the data directory with the server stopped. `inbuxa upgrade` archives a store on the host for you. - The upgrade to layout version 2 is rolling-safe. In a cluster, upgrade the nodes one at a time; nodes still on 27.1-rc.16 keep working meanwhile. - Run with inbuxa-admin 27.1-rc.10 and inbuxa-webmail 27.1-rc.3. ## [2026.11.9.1] - 2026-11-09 Three store tools from 27.1 for the 2026 series, the same as in 27.1-rc.16. No security fixes beyond 2026.11.9. ### Added - Store check (`inbuxa:StoreCheck`). Checks that every blob the server's data holds is in the blob store; a full check also reads each one back and checks its hash. Missing and damaged blobs are reported with the account, the object and the blob id. It runs inside the server, throttled, while mail flows. - Store snapshot (`inbuxa:StoreSnapshot`). Takes a consistent RocksDB checkpoint, plus a filesystem blob store's files if you use one, into `snapshots/` in the store's directory, without stopping the server. The latest two are kept for a backup tool to pick up. - Blob store changes (`inbuxa:BlobStoreChange`). Changing `x:BlobStore` used to strand every blob written under the old setting. Now the server notices the change when it starts and logs a warning naming both stores. A blob it can't find in the new store is read from the old one and kept in the new one, and an administrator can copy the rest over, checked by hash. Going back to an earlier store marks the change reverted. Nothing is deleted from the old store. All three are JMAP objects under `urn:inbuxa:jmap`, for administrators only, with the new permissions `sysStoreCheck`, `sysStoreSnapshot` and `sysBlobStoreChange`. (#340) ### Upgrading - Existing administrator roles are granted the three new permissions once, on the first start. - A store snapshot is refused while a blob store change is unresolved, because it would miss blobs that are only in the old store. - Rolling back to 2026.11.9 is safe: roles keep loading, without the new permissions. ## [27.1-rc.16] - 2026-11-09 27.1-rc.15 with 8 security fixes, 7 bug fixes and three new store tools. ### Security The same 8 fixes as 2026.11.9, below, with the same behavior changes. All versions up to 2026.10.29.1, and 27.1-rc.1 to 27.1-rc.15, are affected. ### Added - Store check: `inbuxa:StoreCheck`, permission `sysStoreCheck`. A quick check confirms that every blob the data links to is in the blob store; a full check also reads each one back and compares its hash. Mail is checked from the message side too, so a message whose blob nothing links to any more is reported. It changes nothing, runs in the background, throttled, and keeps the first 1,000 problems. (#322, #324) - Store snapshot: `inbuxa:StoreSnapshot`, permission `sysStoreSnapshot`. With a RocksDB data store, a consistent snapshot of the store while mail keeps flowing, so a backup tool can copy it instead of stopping the server. A filesystem blob store is copied with it. Snapshots go under `snapshots/` in the RocksDB directory, and the latest two are kept. Other stores are refused, with the reason. (#323, #328) - Changing the blob store: `inbuxa:BlobStoreChange`, permission `sysBlobStoreChange`, event `store.blob-read-through`. A node that starts with `x:BlobStore` pointing somewhere new records the change and keeps the old store's settings. Until it's resolved, a blob the new store doesn't have is read from the old one and kept in the new one. An administrator has one node copy the rest, checked by hash, and what couldn't be copied is reported. Store snapshots are refused while a change is open. (#333, #335) The new permissions are in the superuser and admin grants; tenant administrators are refused. ### Fixed - The four data-loss and DNS fixes of 2026.11.9 (Email/copy, file replace, automatic DNS, BDAT). Automatic DNS here also leaves MX records it didn't create alone. - Deleted accounts: a kept user that never had a password can be restored; the new password becomes its password. (#326, #330) - Domains: deleting a domain removes its queued DKIM, DNS and certificate tasks, and turning one of those to manual removes its task. Turning DKIM back to automatic no longer leaves two tasks. Stale tasks from before end quietly on their next run instead of as Failed. (#327, #332) - Settings: lists stored by a newer release, such as a role with permissions this release doesn't know, load without those values instead of failing. (#329) ### Before you upgrade - Read the behavior changes under 2026.11.9. - Run with inbuxa-admin 27.1-rc.10 and inbuxa-webmail 27.1-rc.3. ## [2026.11.9] - 2026-11-09 8 security fixes from our own review of the server (6 medium, 2 low), and 4 bugs that could lose or damage data. All versions up to and including 2026.10.29.1, and 27.1-rc.1 to 27.1-rc.15, are affected. The same fixes are in 27.1-rc.16. ### Security Medium - IMAP: a mailbox name could inject responses into other users' IMAP sessions (GHSA-8fvp-37f9-h56q) - Live delivery test: the target address could inject SMTP commands (GHSA-489p-848x-h3c9) - JMAP: keyword rights on one mailbox changed messages held in others (GHSA-99x7-xgg9-p2w3) - JMAP: a query with limit 0 returned the whole result set (GHSA-3qw8-p6hx-p3wj) - OAuth: token introspection named the wrong owner and called other tokens active (GHSA-8wvx-v756-mv4h) - CalDAV: sharees saw the details of private events and hidden attendees (GHSA-p9mf-rphv-j27j) Low - WebDAV: with directory queries on, a user could read another tenant's principals (GHSA-84jh-xghf-4m97) - Queue: only the first recipient-domain quota was enforced (GHSA-52gf-846h-j5ch) ### Fixed - JMAP: Email/copy with "destroy original" no longer destroys an original whose copy failed. - Files: replacing an existing file replaces it only once the new write is accepted; a refused write used to destroy it. - DNS: if the DNS provider can't list the records at a name, automatic DNS leaves that name alone instead of overwriting the TXT and CAA records it didn't create. Network errors, timeouts, 5xx answers and rate limits are retried instead of failing the task for good. - SMTP: when a BDAT chunk is refused as too big, the rest of that transaction is refused too, instead of queuing a truncated message. - Roles saved by a newer release, with permissions this release doesn't know, load without those permissions instead of failing, so administrators' roles keep working after a rollback from 27.1. (#331) ### Changed - The store carries inbuxa's own store version. A store marked by a newer release is refused before anything writes to it, with a message pointing at that release or the pre-upgrade backup. (#247) ### Behavior changes Mail - Mailbox names can't contain control characters (tab included) over JMAP, IMAP CREATE/RENAME or Sieve `fileinto :create`. Existing names are left as they are; one with a line break is sent to IMAP clients as a literal. - JMAP: on a shared account, changing a message's keywords needs the right to modify messages on every mailbox the message is in, and `$deleted` also needs the right to remove them. IMAP's `\Deleted` needs only the remove right, as before. - JMAP: `limit: 0` on a query returns no ids, only the total when `calculateTotal` is set. Calendars - CalDAV shows people a calendar is shared with what JMAP shows them: only the time of private and confidential events, and among the attendees of events that hide them only the owners and themselves, unless they may change the event. - People a calendar is shared with can't copy or move a private or confidential event, or a calendar holding one, into another account over CalDAV. - People with write access to a shared calendar can no longer save changes over CalDAV to the owner's private or confidential events (403), since they only see the time of those events. Administration - The live delivery test accepts only a valid email address or domain name as its target; anything else gets 400. A bare domain needs a known public suffix, or a reserved one such as `.test`. - OAuth introspection names the token's owner, and only access tokens are reported active. - WebDAV: with directory queries on, a principal in another tenant answers 403. ## [27.1-rc.15] - 2026-11-08 27.1-rc.14 with one relay addition. ### Added - A relay can use one of the server's certificates. `certificateId` on `inbuxa:Relay` can point at one of the server's own certificates, a wildcard say. The server refuses one that doesn't name the relay, copies it and its key to the relay, and orders nothing for that relay. A renewal is stored as a new certificate, so the relay follows the newest one for the same names. If none will do, the relay keeps its last copy and `certificateError` says why. Setting `certificateId` back to null goes back to ordering. (#316, #318) ### Before you upgrade - Nothing to do. Run with inbuxa-admin 27.1-rc.10, which has the picker for it, and inbuxa-webmail 27.1-rc.3. ## [27.1-rc.14] - 2026-11-08 27.1-rc.13 with two fixes for deleting domains and accounts. ### Fixed - A domain with automatic DKIM can be deleted. If the DKIM keys it generated are the only things still referring to it, they're deleted with it. If anything else still uses it, such as an account, the delete is refused as before and nothing is removed. (#313, #314) - An `objectIsLinked` error has a description, such as "It's still used by 1 x:Account", for clients that show only the description. (#314) - Cancelling a deleted account's destroy task no longer says account recovery isn't supported. With account retention set, a kept account is restored, or destroyed early, with `inbuxa:DeletedAccount/set`, and the message says so. (#313, #314) ### Changed - Management › Directory links to the new Deleted Accounts page in inbuxa-admin 27.1-rc.9. (#315) ## [27.1-rc.13] - 2026-11-07 27.1-rc.12 with one fix. ### Fixed - DNS-01 follows the challenge CNAME. If `_acme-challenge.` is a CNAME, the server follows it (up to eight hops) and writes the TXT record where the chain ends, in that name's own zone. This lets you delegate challenges to a separate zone that is the only one taking dynamic updates. (#310, #311) - A certificate name outside the domain's zone no longer crashes the renewal task. A domain's Origin is used only for names inside it, and other names go to their own registered domain. A record that still can't be placed in its zone fails the order with a reason. (#310, #311) - A task that crashes counts as a failed attempt: it waits the usual retry delay and its error shows on the task. Before, it ran again straight away, which for a certificate renewal meant a new ACME order about once a second. (#311) ### Before you upgrade - Nothing to do. If you set a domain's certificates to manual because of the crash, you can switch them back to automatic. ## [27.1-rc.12] - 2026-11-07 27.1-rc.11 with one fix. ### Fixed - A renewal created through the API replaces the queued one. Creating a certificate renewal task with `x:Task/set`, or a DNS task that renews the certificate afterwards, clears the domain's queued renewals first, as saving the domain already did. A second renewal for the same domain in one request is refused. (#308) ## [27.1-rc.11] - 2026-11-06 27.1-rc.10 with two fixes. ### Fixed - Calendars: a recurring event that starts in UTC keeps its zone. A series whose start is in UTC (a JMAP `timeZone` of `Etc/UTC`) came back in the zone of whoever asked, at a shifted time. Series stored before this are read correctly too, with no reindex. (#303, #305) - Automatic DNS can publish nothing. A domain that only needs its DNS host to answer DNS-01 challenges can leave every record type off, and saving it no longer queues a DNS task that can only fail. (#304, #306) ### Before you upgrade - Nothing to do. If you used the workaround for DNS-01 without records (only TLS-RPT ticked, no report address), you can untick it. ## [27.1-rc.10] - 2026-11-06 27.1-rc.9 with clearer wording for two settings. The server's behavior doesn't change. ### Changed - Additional Hostnames is now Certificate Hostnames, and its description says a list replaces the default names instead of adding to them, what an empty list gets, and how a name without a dot is read. The property and saved configs are unchanged. (#291, #299) - Allow Directory Queries' description says it also decides who a regular user can look up and share with. (#301) - SECURITY.md: vulnerabilities can also be reported privately through the GitHub copy's Security tab. (#300) ## [27.1-rc.9] - 2026-11-06 27.1-rc.8 with certificate and DNS fixes, paging fixes for two query methods, and fuller milter traces. ### Changed - CAA records are off by default. A domain with automatic certificates and automatic DNS used to publish a CAA record at its apex, tied to this server's ACME account, which made another ACME client under the domain fail at its next renewal. New domains leave CAA off. DNS-PERSIST-01 providers publish their `_validation-persist` record whether or not CAA is on. (#294) - Changing a domain's names orders a certificate right away. Editing Certificate Hostnames, the ACME provider or the domain name on a domain already on ACME used to wait for the next renewal. A domain keeps a single queued renewal. (#295) ### Fixed - `x:MaskedEmail/query` and `x:ArchivedItem/query` honor `position`, `anchor` and `limit`; a `position` past the first page returned the first page again. Thanks to Stefan for the masked email fix and its test. (#296, #297) - Milter replies show in delivery traces: when a milter rejects with its own reply, the trace records the SMTP code and text the client was sent. Thanks to Stefan. (#293) ### Before you upgrade - Existing domains keep the records they publish today, CAA included. If another ACME client issues certificates for names under a domain that inbuxa publishes CAA for, untick CAA in that domain's DNS settings, then delete the record at your DNS provider: unticking it doesn't remove a published record. ## [27.1-rc.8] - 2026-11-04 27.1-rc.6 with one security fix and a new journal report format. 27.1-rc.7 was tagged with the security fix, but its build failed and it was never published. ### Security - Allow Directory Queries now does what it says. With the setting off, the default, regular users could still list and look up the other accounts on their server. Listing and searching are now refused over JMAP and WebDAV. A regular user can only look up the accounts they already have a connection to: their own, their groups, people who shared with them, and people on their own shares. Sharing follows the same rule on JMAP, WebDAV and IMAP. (#282, #283) ### Changed - Journal reports follow a public specification, the journal report format (https://coffeylabs.org/specs/journal-report/): a `multipart/report` with a summary for people, an envelope part for programs with one line per recipient, and the message byte for byte. Reports already stored keep the layout they were written in. ### Before you upgrade - On the default, regular users lose directory suggestions in compose, picking anyone they have no connection to in the share dialog, free/busy for people who haven't shared with them, and starting a share with someone new. People already on a share keep it. To keep all of that, turn on Settings › Sharing › Allow Directory Queries. An IMAP connection picks up a change to the setting when it reconnects. - Journal report scripts should read the new envelope part, not the summary, whose wording isn't fixed. The `Bcc` line is now `Not in To or Cc`, and `addressed=not-in-headers` in the envelope part. Reports to an archive declare `BODY=8BITMIME` or `BODY=BINARYMIME`; a server that can't take the message unchanged doesn't get the report, which stays in the built-in journal and shows as an archive failure. - Run with inbuxa-webmail 27.1-rc.3. ## [27.1-rc.7] - 2026-11-02 Tagged with the directory security fix that is in 27.1-rc.8, but its build failed and nothing was published. Use 27.1-rc.8. ## [27.1-rc.6] - 2026-10-31 27.1-rc.5 with one fix. ### Fixed - A new server where a role was made in recovery mode before its first normal start never got the built-in roles, the default Authentication settings or the new-install privacy defaults, so every account was refused with 403. The first normal start now seeds them when there are no roles, or when recovery mode left a note that it ran. A server that has started normally before is unchanged. (#265) ## [27.1-rc.5] - 2026-10-31 inbuxa Relay, and moving a domain in from inbuxa Admin. No security fixes. ### Added - inbuxa Relay: a small relay for a public machine that takes mail in on port 25 for your server and can send its mail on, so the server itself never opens a mail port to the internet. The server dials out to it and manages it under Settings › Mail flow › Sending › Relays. - While the link is up, the server decides on each message during the sender's session: its spam filter, directory and rules answer at the door. - While the link is down, the relay holds mail for your addresses, encrypted to the server's key, and passes it on once the link is back. - Outgoing mail goes through it with DKIM signed by the server, DANE and MTA-STS, REQUIRETLS, and TLS-RPT reports from the relay's TLS results. - The server publishes the relays in its domains' MX and SPF records where it manages DNS, and orders each relay's certificate by DNS-01. - Signed upgrades from inbuxa Admin; a list of who may dial the link; a new link CA without a new join; relay keys wrapped under a key file (`INBUXA_RELAY_KEY_FILE`). - It ships from each server tag as `registry.coffeylabs.org/inbuxa/inbuxa-relay` (linux/amd64, linux/arm64) and as static musl binaries, each with a `.sig` by the inbuxa release key. Run it with the server release of the same version. (#242 to #245, #249, #253 to #274) - Migrations: an administrator can move a domain in from another server, from IMAP, JMAP, Microsoft 365 or Google Workspace, under Management › Directory › Migrations in inbuxa Admin: preview, move, final pass, and the new accounts' passwords once. The server runs the bundled inbuxa-migrate. It needs the new Run migrations permission, which only the Administrator role has. (#275, #276, #278) ### Changed - The store carries inbuxa's own store version. A store marked by a newer release is refused before anything writes to it, with a message pointing at that release or the pre-upgrade backup. (#246) - A store brought over from Stalwart 0.16 has its old web interface switched off the first time inbuxa opens it, so it is no longer served at `/admin`. The log says how to switch it back on. (#251) ## [27.1-rc.4] - 2026-10-29 Fixes for Sieve filters, IMAP pipelining and file deletion. No security fixes. ### Fixed - Sieve: an out-of-office reply (vacation) and `notify` answered mail sent by programs: mail with an `Auto-Submitted` header other than `no`, other autoresponders included (RFC 3834). They now stay silent for it. (#235) - Sieve: `string` depended on the order of its keys; a test is now true when any key matches. (#236) - Sieve: `date :regex` and `currentdate :regex` compared the date as a wildcard pattern, so a regular expression never matched. (#236) - IMAP: a `SELECT` pipelined with the `FETCH` or `STORE` after it got `BAD No mailbox is selected` for those commands. Each command is now checked when it runs. A `FETCH` pipelined after `UNSELECT`, `CLOSE` or `UNAUTHENTICATE` is refused with BAD; before, it closed the connection. (#237) - Files: the task that keeps a deleted file for undelete could wait up to five minutes on a quiet server. Deleting a file now starts it at once. (#234) ## [2026.10.29.1] - 2026-10-29 7 security fixes from our own review of the server, all medium. All versions up to and including 2026.10.29, and 27.1-rc.1 to 27.1-rc.2, are affected. The same fixes are in 27.1-rc.3. ### Security - OpenID Connect directories: tokens issued to other applications could sign in (GHSA-v3gv-4j87-3mg2) - OpenID Connect directories: opaque tokens from any application could sign in (GHSA-2p7h-9hv3-4292) - Calendars: one calendar object could cost tens of seconds of CPU to expand (GHSA-pw47-x75c-v45r) - IMAP: one session could start an unbounded number of operations at once (GHSA-6x5j-v349-9v79) - Sieve: a script's CPU limit did not count the message it scanned (GHSA-whf3-4g3x-w7r6) - Sign-in: bearer tokens were offered and accepted on unencrypted connections (GHSA-5phc-xg42-vp24) - OpenID Connect directories: provider responses were read without a size limit (GHSA-6hcw-cw8m-jcx7) ### Behavior changes Sign-in and tokens - OAuth bearer sign-in (OAUTHBEARER, XOAUTH2) always needs TLS in IMAP, POP3, ManageSieve and SMTP submission, even with plain-text sign-in allowed. A proxy that terminates TLS has to report it through the proxy protocol. Password sign-in is unchanged. - New OpenID Connect directories must set a required audience (the client ID registered for this server). Existing directories without one keep working, log a warning each time they load, and appear on the Security to-do list; a later release will require it. An audience, once set, can't be removed. - OpenID Connect directories have an opaque-token setting: reject, check by token introspection (RFC 7662, with a client ID and secret), or accept on the user info endpoint's word. New directories reject opaque tokens unless an introspection endpoint is set. Existing directories keep accepting them through user info; set the policy to close the hole. - Responses from an OpenID Connect provider larger than 1 MiB are refused. A token without a key ID is checked against freshly fetched provider keys when the cached ones don't verify it. Mail and calendars - IMAP: a session runs at most four read operations at once, and commands that change mail or the session run one at a time in the order sent. A NOOP or LOGOUT sent during a long FETCH answers after it. - Sieve: the CPU limit (`max_cpu_cycles`) also counts the message bytes that tests scan: every 64 KiB compared, times the number of keys, costs one instruction, so a script can scan about 300 MiB per message by default. A script that goes over stops with the usual CPU-limit error and the message is kept. - Calendars: recurrence expansion has one work limit per calendar object (about 50 ms). Objects whose rules use it up are stored with the instances found so far, and the server logs a calendar rule expansion error. ## [27.1-rc.3] - 2026-10-29 ### Security The same 7 fixes as 2026.10.29.1, above, with the same behavior changes. ### Added - `/.well-known/source`, as https://coffeylabs.org/specs/well-known-source/ defines: the server's name, version and license, its public repository, and the commit the build was made from. A build without a full commit id answers 404. (#227) ## [27.1-rc.2] - 2026-10-29 ### Security The same 34 fixes as 2026.10.29, below, with the same behavior changes. ### Changed - Spam filter: the rules stay on v3.0.2 and are maintained in this repository from now on, versioned `3.0.2-inbuxa.N`. Later upstream rule releases rename tags and settings this server doesn't use, and would score the AI classifier's verdicts at 0 on new installs. The first change adds scores for the DKIM2 result tags the server already emits. (#225) ## [2026.10.29] - 2026-10-29 34 security fixes from our own review of the server: 2 high, 31 medium and 1 low. All versions up to and including 2026.10.28, and 27.1-rc.1, are affected. The same fixes are in 27.1-rc.2. ### Security High - SMTP: BDAT reserved the declared chunk size before any data arrived (GHSA-v35h-m53h-jxgx) - IMAP and ManageSieve: unauthenticated clients could make the server hold about 100 MB per connection (GHSA-fvgm-5jfp-rpmm) Medium - OpenID Connect directory: an ID token from another application could sign in as its user (GHSA-mfq6-g7cj-pxp7) - OAuth: a refresh token could be redeemed without its client's secret (GHSA-fw6c-mjh4-5mf5) - JMAP: a property listed many times in /get or Email/parse was rendered once per mention (GHSA-rj6m-4hv5-r34v) - Calendar: a deeply nested iCalendar component could crash the server (GHSA-w697-j727-2ffc) - Sieve: regex tests could cost seconds of CPU per delivery and 10 MiB each (GHSA-7m5r-6wf6-c55c) - Sign-in: password hash checks could stall every protocol (GHSA-h6mp-735j-h3c7) - SMTP: TLS reports and MTA-STS policy fetches could be aimed at the server's own network (GHSA-pf39-vc2q-8r54) - JMAP: an open WebSocket ignored token expiry, account locks, password changes and removed shares (GHSA-wf4x-2gmj-c46w) - WebDAV: one lock token let a request through every other lock it touched (GHSA-65q8-w3r7-3v6q) - WebDAV: file COPY and MOVE skipped the add and delete rights (GHSA-qpp4-998x-cg8w) - WebDAV: a copied folder kept its own sharing inside someone else's folder (GHSA-6593-48rg-fhgq) - POP3: OAuth bearer tokens were accepted over clear text with plain-text sign-in turned off (GHSA-vh5g-wvxj-x8gf) - Sign-in: an address whose ban had ended was never banned again (GHSA-9pmc-86q4-q82f) - Spam filter: one message could make the server issue hundreds of URL lookups (GHSA-j825-8cpr-r2w3) - SMTP: a From header naming two domains skipped DMARC (GHSA-f5r6-c8px-rwmm) - Mail: many messages with one subject made every later delivery slower (GHSA-v2cm-f2pc-x8rv) - JMAP: web push could be aimed at the server's own network (GHSA-w27c-h3jv-85c9) - HTTP: bearer tokens kept working after they expired (GHSA-hhcw-frq4-p343) - SMTP: an on-path attacker could switch off MTA-STS for a domain (GHSA-w5xh-2264-rh54) - Calendar: iTIP REFRESH and REPLY were accepted from people who weren't attendees (GHSA-wjcq-c7g4-q568) - IMAP: a JMAP keyword could inject responses into other users' IMAP sessions (GHSA-xprp-24v8-28c8) - IMAP: a user with flag rights on a shared mailbox could set `\Deleted` (GHSA-2qxg-g2qm-983v) - WebDAV: dead properties and lock owners could inject XML into other users' responses (GHSA-rfcv-29qm-49qj) - IMAP: a revoked mailbox share stayed readable for as long as the connection stayed open (GHSA-qq56-9g55-5gh2) - Calendar: a local user could send iMIP invitations and replies in another person's name (GHSA-9m6q-mwc8-34c4) - Sharing: sharing a calendar or address book exposed an unrelated event or contact (GHSA-8fv8-q3vx-c544) - JMAP calendars: sharees saw the details of private and confidential events, and hidden attendees (GHSA-vw7g-hjcp-wcch) - Calendar: invitations to an event that hides its attendees listed every attendee (GHSA-cr6x-q56w-r597) - WebDAV: sync-collection showed a sharee what was deleted from the owner's other collections (GHSA-xjqh-p8qf-m4ff) - WebDAV: nresults 0 lifted the cap on query results (GHSA-hgf7-8wq9-hff3) - Quota: calendar, contact and file writes could go past the disk quota (GHSA-vmw3-q5hr-6p5f) Low - JMAP: a message with two Bcc headers, or a very long one, sent the Bcc addresses to every recipient (GHSA-gh3w-qwv3-57cj) ### Behavior changes Sign-in and tokens - External OpenID Connect directories: a JWT the provider marks as an ID, refresh, offline or logout token (Keycloak `typ`, Amazon Cognito `token_use`) is refused as a bearer token. Clients that were sending ID tokens instead of access tokens stop signing in. - OAuth refresh tokens can be redeemed only by the client they were issued to, and a confidential client has to authenticate to do so. A request naming another client gets `invalid_grant`. Refresh tokens issued before tokens named their client go to the first client that authenticates with them. - Bearer tokens over HTTP stop working when they expire. Before, a token could keep working for up to the access-token lifetime after its first use. - A JMAP WebSocket is closed with an `unauthorized` error when its token expires, the account is locked or deleted, the password changes, or the app password or API key it used is removed. Requests on a WebSocket count toward the per-account request limits. - POP3 with plain-text sign-in off: OAUTHBEARER and XOAUTH2 are refused on clear-text connections, and CAPA lists no SASL mechanism there. - IMAP and ManageSieve accept at most 64 KiB per request before sign-in. - Slow password hashes (bcrypt, crypt family, Argon2, PBKDF2, scrypt) are checked at most one per CPU at a time; under a sign-in flood the rest wait their turn. - An address that fails sign-in again after its ban ended is banned again. Mail - DMARC: a From header with addresses in several domains is checked for each domain, and the strictest result applies. Some mail that was accepted before will now be rejected or quarantined. - MTA-STS: a cached policy stays in force until it expires when a new one can't be fetched. A domain whose `mta-sts.` host resolves only to private addresses fails its policy fetch; in the default optional mode mail still goes out, but with MTA-STS set to "require" that domain is deferred. - TLS reports over https connect only to public addresses and follow no redirects. A report address that is a local IP literal is skipped. - Spam filter: at most ten URL-redirector lookups per message, over https with the certificate checked. - Threading compares a new message with the newest 512 messages that share its subject. A reply whose thread has no message among those starts a new thread. - iMIP: a scheduling message from a signed-in sender is credited to its envelope sender, not its From header. From counts only when it passed DMARC. - JMAP: keywords must be valid names (1-255 printable ASCII characters, none of `( ) { ] % * " \`). Email/set and Email/copy refuse others with `invalidProperties`, and Email/import refuses the whole message. IMAP leaves stored keywords that aren't atoms out of FLAGS. - IMAP: on a shared mailbox, setting or clearing `\Deleted` needs the remove right ("t"), and other flags the modify right ("w"), on every mailbox the message is in. Flags the user can't change are left alone. - IMAP: once a share is removed, commands on that still-selected mailbox get `NO [NOPERM]`. CLOSE and UNSELECT still work. Calendars, contacts and files - JMAP calendar privacy applies to everyone but the calendar's owner: private, confidential and secret events show sharees only their time, and events that hide attendees list only their owners and the reader. - Invitations, updates and cancellations for an event that hides its attendees go out as one message per recipient. - iTIP REFRESH is answered only for attendees of the event. A REPLY for an instance the organizer didn't change is stored only from an attendee of the series and only for a date of the series. - Components nested more than 32 levels deep are dropped when an event is converted for JMAP. - Quotas: calendar events, contacts and files count against the quota of the account they are written into, not the writer's, and files written through JMAP now count. Accounts near their quota may start getting `overQuota`. - WebDAV file COPY, and MOVE from another account, takes the destination folder's sharing, as an upload does (RFC 3744). COPY or MOVE into a folder needs the right to add to it; overwriting needs the right to delete every node replaced. - WebDAV writes need a lock token for every locked resource they touch. Holders of a shared lock still need only their own token. - WebDAV sync-collection reports only deletions inside the synced collection. - WebDAV `0` is treated as 1. Sieve and push - A Sieve `:regex` pattern that compiles to more than 1 MiB is refused ("Invalid regular expression"). With Unicode `\w`, `\w{50}` is too large and `\w{20}` fits. Stored scripts with such a pattern stop matching that test. - Web push goes only to public addresses and never through a configured proxy. Push to a host name on your LAN (a self-hosted ntfy or UnifiedPush server, say) no longer works. ## [27.1-rc.1] - 2026-10-28 The first release candidate of the 27.1 series: 2026.10.28 under the new version scheme. ### Changed - One version for the whole workspace, in Cargo.toml, shown as tags name it: `27.1`, `27.1.3`, `27.1-rc.1`. Release candidates are images `:27.1-rc.N` and `:27.1-rc`, never `:latest`, which stays on the 2026 series until 27.1 ships. (#223, #224) ## [2026.10.28] - 2026-10-28 A security release: 21 fixes, among them several that let a sharee or a limited administrator reach data or accounts they shouldn't, and several ways one request or one inbound message could tie up or crash the server. (#222) ### Changed - An unknown JMAP capability is answered as `unknownCapability`, not `notRequest` (RFC 8620 §3.6.1). (#210) - IMAP: OBJECTID+ follows draft rev 08. LIST takes the OBJECTID return option, and SELECT/EXAMINE by identifier never matches a mailbox the user may not read; it falls back to the name instead. (#219) - Sieve scripts that require the extension names from before the rename (`vnd.stalwart.expressions`, `vnd.stalwart.while`) can be saved again after an edit. Only the inbuxa names are advertised. (#212) ### Before you upgrade - Behind a proxy? `X-Forwarded-For` and `Forwarded` are read only from trusted proxies: loopback and private addresses by default. If your proxy connects from a public address, list it in `INBUXA_HTTP_TRUSTED_PROXIES`. - PKCE is S256 only (since 2026.10.19). The console signs in with S256 over HTTPS or on localhost; on plain `http://` from another machine the browser has no Web Crypto and sign-in fails, so serve the console over HTTPS. - Device sign-ins (OAuth device grant) in progress during the upgrade have to start again. ## [2026.10.19] - 2026-10-19 ### Security - OAuth: only S256 PKCE challenges are accepted. The authorization endpoint took the plain method, and assumed it when a challenge came without one, though the metadata listed only S256. Codes issued with plain before the upgrade still verify. (#205) - A registry write that sent back a stored secret as read (`****`) stored `****` as the secret: a webhook's password or signing key, a store's or directory's password, a DNS provider's key, an OAuth client's secret. Written straight at a secret it now keeps the stored one; anywhere else the write is refused with `invalidProperties`. (#206) ### Fixed The last upstream import, 0.16.27 (#204): - IMAP: `APPEND` and `COPY`/`MOVE` into another account's mailbox did not change the JMAP `EmailDelivery` state. - IMAP: `SEARCH RETURN` with `MIN` or `MAX` alongside `ALL` returns only the minimum and maximum in `ALL`, and `SAVE` combined with `MIN` or `MAX` and `ALL` or `COUNT` saves only the minimum and maximum. - iTIP: merging a received update didn't copy the `VTIMEZONE` definitions it carried, so an occurrence moved by an external organizer could reference a `TZID` the stored event no longer defined. - OAuth: loopback redirect URIs without a path (`http://127.0.0.1:45678`) weren't matched against registered loopback URIs on any port, and were rejected by dynamic client registration. - WebDAV: `PROPFIND` `allprop` returns `DAV:current-user-principal`, so Windows Explorer shows every file and folder under the account name. ## Earlier releases The dated releases from 2026.9.20 to 2026.10.16 list their changes on their release pages. Until 2026.10.19 this file carried the changelog of the codebase inbuxa was forked from, up to 0.16.27, the last release imported. It is in the history: `git show v2026.10.19:CHANGELOG.md`.