What's Changed
Added
- A sign-in refused for want of a role says what was read. The audit entry and the warning of a
sso_unmappedorsso_no_default_rolerefusal give the group values read (at most 20) and their source (ID token, UserInfo, SAML attributes, LDAP search base) and, when none was read, the names of the claims or attributes that may carry them; the LDAP mapping test returns the same text, and the person refused is told no more than before. - SAML group attribute shortcuts. Buttons under a SAML provider's attribute mapping add a
→ groupsrow forRole,member(Keycloak),memberOforurn:oid:1.3.6.1.4.1.5923.1.5.1.1(isMemberOf, Shibboleth); nothing new is read until such a row is saved. The mapping editor of a SAML or OIDC provider names the attributes of oldergroups → attributerows that no row reads any more once such a row exists. - Test an OIDC or SAML sign-in before enabling a provider. Test sign-in on the provider screen signs in at the IdP in a new window and shows the identity, groups, verdict, role and account a sign-in would give, through the login's own checks, without signing anyone in or changing any account; the test is recorded as
sso_sign_in_test, never as a login. - An account may be opened by several SSO providers. An administrator attaches providers to an account one by one (
/api/v2/users/{id}/sso-identities, the account panel of Users), each with its own stable identity, resets the identity of one or detaches one; a sign-in never attaches a provider to an existing account by itself. - SSH CAs on an HSM key. An SSH user or host CA can be created on an RSA or NIST ECDSA key of an HSM provider (
hsm_key_id, withread:hsm), its private key never leaving the token. An HSM key backs one authority, X.509 or SSH: the screens offer only unused keys and show the server's refusals (#445). - Sync keys on the HSM page. The button reads a provider's keys again from the token, one sync at a time, and its message counts the keys the token could not read and the missing keys it kept; a slot re-keyed with another algorithm no longer keeps the old public key (#444).
- OpenSC in the Docker images. The standard and FIPS images ship OpenSC and the PC/SC client library, so a smart card reader shared from the host can serve as a PKCS#11 provider (see docs/HSM_DOCKER.md).
- A restore never gives one HSM key to two authorities. A restore that would leave one HSM key backing two authorities, X.509 or SSH (an archive that holds them so, or a merge onto a key already held here), is refused with both names and changes nothing (see UPGRADE.md).
- Russian interface language, the tenth; the in-app help stays in English for it.
- The Helm chart is published. Each stable release pushes the chart to
oci://ghcr.io/neyslim/charts/ucm, versioned as the release in semver form (UCM 3.0 is chart3.0.0, where the chart in the source tree said0.5.0), itsappVersionthe image tag of that release. - An SBOM for every deliverable. Each release publishes a CycloneDX 1.6 SBOM per package file and per image platform (standard and FIPS, amd64 and arm64), bound to its file hash or image digest and listed in
sbom-index.json; no image is tagged and no release is created without all of them (see docs/SBOM.md). The fonts of the web interface now ship with their SIL Open Font License texts, named inNOTICE. - A pending Microsoft CA request completes from UCM. A request approved on the Windows CA console is finished by Check status in the CSR details or by the inventory sync (
completedanddeniedin its answer), instead of staying Pending until approved from the CA Control Panel, andGET /api/v2/microsoft-cas/{id}/requests/{request}requireswrite:certificatesinstead ofread:certificates. A request whose CSR receives a certificate from another source (a local CA, an import) is closed assuperseded, a direct signing beaten by one answers 409, and a certificate the Microsoft CA still issues for it enters as a separate certificate only through a CA Control Panel approval or a full inventory sync (full: true). - Restore from the screen, any size, with a report. The restore picker accepts archives up to the server's ceilings instead of 10 MB and shows the upload progress; after a restore with warnings, a report with a per-element To do list stays open until closed, and an archive without data says that nothing was restored (
data.todoin the restore answer,restore_limitsin the backup listing). An archive written before 3.0 aboveBACKUP_MAX_UPLOAD_MBis refused before the upload, and a 413 from a proxy no longer reads as a network error. - Unique subject per CA. A CA option, off by default, refuses a certificate whose subject a certificate of another entity already carries under that CA, expired or revoked included (RFC 5280 §4.1.2.6), on every issuance path; the same key, a renewal, the ACME account or the Windows SID make the same entity, and manual issuance and CSR signing let an operator attest, audited, that a new certificate is for the entity already carrying the subject.
- Size ceiling for the backup archives. Settings > Backup shows the space the stored archives take and sets a ceiling in MiB (
max_total_mbonPATCH /api/v2/settings/backup/schedule, 0 for none,backup_max_total_mbin General settings); retention removes the oldest archives over it, audits each removal by name, and reports a total it cannot bring under it without removing a protected archive; a manual backup removes nothing and its response carries the samewarning. - 3.0 upgrades a database last started by 2.254. Install 2.254 and start it once before installing 3.0: a database whose registry lacks a migration of 2.254, or has no registry, is refused at startup with the number of missing migrations, and nothing in it is changed (see UPGRADE.md). A registry naming a migration that neither 2.254 nor this release knows, recorded by a later release or by hand, is refused the same way.
- The DEB and RPM packages refuse to install over a UCM older than 2.254. The pre-installation script stops before any file is written, names the installed version and says to install 2.254, start it once, then install the package; the installed version keeps running. The packages no longer run the 1.x to 2.0 upgrade, which goes through 2.254.
- A new package install archives what a removed older UCM left. When 3.0 cannot start on the SQLite database a removed UCM left, the DEB and RPM move that installation's data and configuration, unchanged, into
/opt/ucm-archive/<date>/(root only, with restore steps; what cannot be renamed out of its filesystem goes into a root-only.ucm-archive-<date>/beside it) and start a new, empty UCM; a DEB removed but not purged is no longer refused. - A copy of the SQLite database before the upgrade from 2.254. The first 3.0 start copies a 2.254 SQLite database to
ucm.db.2.254-before-3.0next to it, checks the copy, and does not start the upgrade without it; a start after an upgrade cut short keeps that copy, never copies the half-upgraded database in its place, and resumes without a copy, saying so, when that copy is gone. PostgreSQL gets no copy, take apg_dumpfirst (see UPGRADE.md). - A parent whose signature cannot be verified is confirmed by hand. A CA left without a parent because UCM cannot verify the parent's signature (SLH-DSA, GOST) now names its probable parent in every import response and in its details, and
POST /api/v2/cas/{id}/parent-confirmation(write:cas,"confirm": true) links it after checking everything again, audited with who and when;DELETEon the same route withdraws it. - Common name pattern for Windows enrollment (#442). A template whose subject UCM builds can set the CN from fixed text and directory variables (
{hostname}-WiFi,{username},{displayName}), alone or followed by the directory path; the SAN and the SID extension are unchanged, and a request whose directory object leaves a variable empty or gives a CN over 64 characters is refused. A subject the pattern builds is refused at issuance, renewal and approval when a certificate of the same CA, whatever its status, carries it (compared per RFC 5280, case and extra spaces ignored) without the requester's SID: a secondJohn Smithunder{displayName}is refused, the first one renews. - Several Active Directory connectors.
/api/v2/ad-connectorsmanages one connector per forest or domain, and each lookup, renewal check and health probe uses the connector of the domain concerned. A SID the directory returns outside its connector's domain is refused at issuance. - Windows enrollment profiles.
/api/v2/windows-profileslists (paginated, searchable), creates, changes, disables and deletes the profiles; the default profile and a profile with valid issued certificates cannot be deleted. - Windows profiles and AD connectors in Settings. The Windows Autoenrollment tab lists the profiles with an editor and a checklist per profile, and the Active Directory tab lists the connectors with their health and domain SID. The single-row routes
/api/v2/xcep/config,/wstep/config,/kerberos/config,/kerberos/keytaband/ad-connectorare removed (see UPGRADE.md). - A certificate names its Windows profile. The certificate list and detail carry
windows_profile_idandwindows_profile_name(null once the profile is deleted), and the detail panel shows it. - UsernamePassword accounts only on the default profile.
xcep_username,xcep_password,wstep_usernameandwstep_passwordset on another profile are refused (400) instead of saved and never read; the editor shows them for the default profile only. - WSTEP says why it cannot build a user subject. A user missing from the directory and an account without
userPrincipalNameare refused with their own message (and audit reason for Kerberos) instead of "could not resolve the requester"; who is served is unchanged. - Default Windows profile from the editor. While no profile is the default, the profile editor offers Default profile (new or saved profile) and the API accepts
is_default: trueon creation and update; once one exists, another designation is refused with 409 naming it. - The group refusal names what required it. A requester refused for the AD group set on the profile is told so ("this profile"), not "this template".
- The connector domain refusal offers only exits that exist. A profile with a keytab is told to take another connector (or none) instead of changing a realm its keytab fixes; the guide no longer says a keytab can be deleted.
kerberos_refusalreports a profile with no keytab. A profile with a realm but no keytab gavenullwhile its clients got a 401; it now saysno key for realm ..., in the list and the detail alike.- Windows enrollment guide. The refusal table no longer lists an unreachable case and says a realm without key leaves no audit entry; renewal says a refused TLS handshake leaves none either and that a profile outside the mTLS CA has only Kerberos; the setup steps list the prerequisites (trusted HTTPS certificate, restart on the WSTEP switch, machine templates).
- Connector domain and profile realm stay together. Changing the domain of a connector moves the realm of its keytab-less profiles that held the old domain; a profile with no realm or already on the new domain is left alone, and any other refuses the save (409) with its name. A profile whose realm differs from its connector's domain is flagged (
connector_domain_mismatch). - No default admin password. The first administrator's password is provided (
INITIAL_ADMIN_PASSWORD_FILEorINITIAL_ADMIN_PASSWORD) or generated intoDATA_DIR/initial-admin-password(0600, never logged), which must be changed at first login and is deleted by that change; packages, image and chart no longer carrychangeme123. - RPM for RHEL, AlmaLinux and Rocky Linux 10 (
ucm-el10-3.0-rc1-1.el10.noarch.rpm); the updater picks it on an EL10 host, and an EL host with no matching build is still offered the Fedora RPM, which needs a systempython3of 3.12 or later. - Inspector: the detail of a certificate, CA, trusted certificate or user certificate opens in a tabbed dock at the bottom or right of the window (full screen on mobile), many objects at a time, instead of floating windows; its position and size follow the user and an invalid value is refused with a 400.
- Top bar: a theme menu and a notification bell fed by the dashboard queue and the audit log (read marks saved per user); after sign-in you return to the page that asked for it, and a server rate limit is shown on the login page.
- SSO: an OAuth2/OIDC or SAML sign-in also returns to the page that asked for it; the server keeps that page with the login (never sent to the IdP), uses it once and checks it again.
- Sessions: on PostgreSQL without Redis, server-side sessions are stored in the database so a session opened on one node is valid on the others; expired rows are purged by the scheduler and every row is deleted after a restore.
- Session key: with encryption enabled (in HA, the KEK every node holds), Settings > Security rotates
SECRET_KEYinto the database under the data key; every node loads it without a restart and the previous key stays accepted for 24 hours, so nobody is logged out. Without encryption the rotation still writesucm.envand restarts, and is refused in HA. - Dashboard:
GET /api/v2/dashboard/queue,/health,/complianceand/issuance(scoperead:dashboard, now held by every built-in role and grantable to groups) return the work queue, per-protocol health, compliance counters and issuances and in-place renewals per CA, computed in SQL without parsing certificates. A repeatedkindparameter restricts the queue items to those kinds. Certificates gainkey_bitsandsig_hashcolumns, backfilled at the upgrade. - Overview: the home page is now six independent cards (work queue, automation health, hierarchy, expiry horizon, compliance, latest operations) with layout presets saved per user; each compliance counter and renewal kind opens the certificate list filtered to exactly what it counts, shown as removable chips. The list accepts the matching filters and refuses an invalid filter value with 400.
- Encryption: private keys are encrypted under a data key stored in the database and wrapped by
master.keyorKEY_ENCRYPTION_KEY, which stays the only secret to back up; the data key can be rotated from Settings > Security without changing it, and the key is changed by hand with the previous one beside it. - SSO: an LDAP provider takes an optional group search base (
ldap_group_base_dn), the only base of thememberoruniqueMembergroup search when set. - Encryption: after a data key rotation or a disable, the old data keys stay readable for seven days and are dropped only once a daily sweep finds nothing left under them, so a value committed late under one cannot become unreadable; the next rotation waits only until every node has reloaded, then two minutes (ten at most).
- Renewal: a deploy binding or the HTTPS binding that an EST, SCEP or WSTEP renewal leaves on the previous certificate (no key in UCM) is now flagged in the API (
needs_attention), the bell, the overview queue and acertificate.binding_left_behindevent; it can be reattached once UCM holds the key, and importing that key moves it. - The overview follows the data. Its cards reload on their own, without leaving the page, when certificates or CAs change, when the live link comes back and when the tab returns, and a Refresh button reloads them all; the health card shows the TSA signer and, with the right to read settings, whether email, webhooks and auto-renewal are set up, the work queue lists trust store certificates that expire within 90 days or have expired (the expired ones after every other critical item), and the expiry horizon marks the TSA signer and delegated OCSP responders.
High availability
- One database key for the cluster. A node with
UCM_HA=1requiresUCM_DB_ENCRYPTION_KEYand never uses the key derived from its machine, so integration secrets saved on one node work on every other; the chart and the Compose recipe provide it, andpython -m scripts.rewrap_db_keyre-encrypts a single instance's secrets under it. - HA mode.
UCM_HA=1runs UCM on several nodes sharing one PostgreSQL database: a node requiresSECRET_KEYandKEY_ENCRYPTION_KEY, refuses operations that only change the local host, and reportsha,node_idandleaderin/health. The Helm chart probes readiness on/health/ready, where safe mode answers 503 only withUCM_HA=1, and liveness on/health/live. - One scheduler, elected. One node runs the scheduled tasks, elected by a PostgreSQL advisory lock and replaced within seconds when its connection drops, or about a minute after its host vanishes. Nodes heartbeat in
cluster_nodes, "run now" works from any node, and startup migrations run on one node at a time. - Real-time events reach every node. Without Redis, an event raised on any node reaches the browsers connected to every node, once each. On any PostgreSQL install without Redis, a
LISTEN/NOTIFYwake-up delivers events within a second. - Rolling package update. On DEB/RPM, an update installed from the web interface or by the automatic install updates the nodes one at a time, each from a SHA-256 verified download of its own package family, and stops at a node not back on the new version until an admin resumes it; each node, leader or not, reports the result of the update it installed. Docker and Kubernetes refuse it: the orchestrator updates the image.
- Settings reach every node. Rate limits, syslog forwarding and the admin allowlist changed on one node take effect on every node within seconds, and a service restart restarts the nodes one at a time, containers included.
GET /api/v2/system/service/statuslists every node. - Encryption from any node. Enabling, disabling and rotating encryption work from any node and reach the others within seconds. A node whose key-encryption key fingerprint differs from a live peer's refuses to start unless the database names its key.
- HTTPS certificate and keytab on every node. The HTTPS certificate can be applied or regenerated on any node, and every node writes it from the database before its TLS socket opens. An uploaded Kerberos keytab is stored in the database and written on every node.
- Scheduled work runs once. Expiry e-mails and webhooks, the scheduled CRL of a CA, Microsoft CA CRL revocation announcements, scheduled reports and scheduled backups go out or are written once when two nodes run them during a leader change, and two nodes serving orders that share one upstream ACME proxy challenge start its dns-01 automation once.
GET /api/v2/settings/backup/schedulereportsin_progresswhile a scheduled backup is being written, and a restore keeps the target's own last run and reservation. - Restored files on every node. After a restore on one node, every node writes and removes its own certificate and key files to match the database at its turn of the rolling restart, and serves the HTTPS certificate of the archive from then on, as it does a regenerated one.
- Backup restore in HA. A restore runs from any node: every node first acknowledges a cluster maintenance, during which it refuses writes with 503 and
Retry-Afterand leaves/health/ready, then the restore commits or rolls back as on a single node and each node serves the restored state after its turn of the rolling restart (see docs/HA.md)./api/v2/system/maintenancesets, shows and lifts the same maintenance by hand around a restore with PostgreSQL's own tools. - HA backups are shared. In HA, scheduled and manual backup archives are stored in PostgreSQL: every node lists the same ones with the node that wrote them, downloads any of them byte for byte even once that node is gone, and retention runs once, on the leader. Settings > Backup says that these archives are lost with the database and are to be downloaded out of the cluster (see docs/HA.md).
- No SoftHSM token in HA mode. With
UCM_HAthe image no longer creates its default token, and creating a SoftHSM provider, pointing one at SoftHSM, generating or syncing its keys, or creating a CA on a SoftHSM or local smart card reader key (OpenSC, YubiKey PIV) answers409ha_mode. A SoftHSM provider already in the database is kept, named in a warning at start, and keeps signing on the node that holds its token. - Two nodes cannot share a node identifier. A node started with the
UCM_NODE_IDof a live node (two cloned VMs keep the same host name) refuses to start with a message naming it, and a running node whose identifier another process took fails/health/readyand stops leading instead of overwriting its row; a node restarted after a crash waits at most three minutes for its previous heartbeat to expire. The Helm chart in HA mode has a startup probe covering that wait, so the liveness probe does not restart such a container before it serves. - A rolling update refuses a release that requires every node stopped. A migration that removes or narrows what the release before uses now carries a full-stop mark, published with the release: the rolling update refuses it before installing anything, and in HA a node refuses to apply it while a node of another version still runs. The upgrade from 2.x to 3.0 is such a full stop.
- A node whose clock runs ahead no longer repeats another node's work. ACME order leases, the scheduled backup reservation, webhook and deployment delivery claims, notification deduplication and discovery schedules are written and compared on the PostgreSQL server's clock, so a renewal, backup, delivery, mail or scan is not run twice when node clocks drift apart.
- The application reconnects after a PostgreSQL restart. Every connection taken from the pool is checked first, so the first request after the database comes back no longer fails, and TCP keepalives with a 10 second connect timeout, and a 30 second TCP user timeout, make a partition fail requests instead of leaving them hanging.
- HA packaging. The Helm chart takes
ha.enabled(rolling update, PodDisruptionBudget, sharedKEY_ENCRYPTION_KEY),docker-compose.ha.ymlstarts two nodes behind a balancer, the image entrypoint skips SQLite in HA mode, anddocs/HA.mddocuments the deployments. A render without cluster access (helm template, GitOps) refusesha.enableduntil both keys are set, rather than write new ones at each sync.
Changed
- SQLite planner statistics. A scheduled task refreshes them every hour, and Optimize Database now reads a bounded sample of each index (
PRAGMA analysis_limit = 1000) instead of every row. - Manual backups run in the background. Creating a backup no longer holds the web worker: the dialog shows its progress, can cancel it and offers the archive once it is ready, a reload picks it up again, and past 25 seconds the create routes answer
202with a job to follow; a shorter one answers as before. - The audit chain verification runs in the background. It no longer holds the web worker: past 25 seconds
GET /api/v2/audit/verifyanswers202with a job to follow (/api/v2/system/jobs), the screen shows its progress and can cancel it, andPOST /api/v2/audit/verifystarts one explicitly; a shorter one answers as before. During an HA cluster maintenance no job starts: both answer503withRetry-After. - Restores run in the background. A restore no longer holds the web worker: its dialog shows its progress and can cancel it until it commits, the sign-out that ends it asks to sign in again, and the report shows after that sign-in; past 25 seconds both restore routes answer
202with a job to follow, a shorter one answers as before. - Writes are refused during a restore on a single node too. While a restore runs, every write answers
503withRetry-Afterand scheduled tasks wait; reads, OCSP, CRLs and health checks are served, as in HA. UCM_WORKER_TIMEOUT. The gunicorn worker timeout, 120 seconds until now, is set by this variable from 30 to 3600 seconds; a value outside these bounds stops the start with a message naming it and exit code 78, which systemd does not restart.- PostgreSQL databases must be UTF8. A database in another encoding (
SQL_ASCII) is refused at startup with a message naming its encoding, where it failed withTypeError: cannot use a string pattern on a bytes-like object. - One systemd unit for both packages; a refused database or a failed upgrade is not restarted. Either one stops the service with exit code 78, which the unit leaves failed until you act: the DEB unit restarted it three times, the RPM unit every 10 seconds without end. A failure while it runs restarts it, at most 5 times in 5 minutes; a restart asked from the interface, an update or a package install is not counted.
- The Docker image checks the disk before a SQLite start. Before its copy of the database it refuses to start, with the space needed and available, when the backups directory lacks the database's size for that copy (
UCM_BACKUP_ENABLED=falseskips it) or, on a start that upgrades the database, its filesystem lacks that size again; a start on a database already up to date only needs room for the copy. - The Helm chart holds its liveness probe until UCM answers. A startup probe on
/health/live(startupProbe, 360 s by default) now comes by default, with or without HA, so a long first start that upgrades the database is no longer restarted after about 100 seconds. - The start log times each migration. Each migration and each step of the 3.0 upgrade logs its start, its end and its duration, and a step that reads certificates in batches logs its progress at most every 30 seconds.
- TSA requires a dedicated signer by default. Without a dedicated signing certificate,
/tsano longer signs with the CA certificate (RFC 3161 §2.3) unlessrequire_dedicated_certis turned off, which an upgrade from 2.x does for a TSA that signed with its CA, flagged as a warning on the overview. In that default mode the signing CA only issues the dedicated signer, so a CA whose key cannot sign tokens (EdDSA, or ML-DSA in FIPS mode when the module does not approve it) can be chosen. - Docker images: the
latestandlatest-fipstags now go to the highest stable version published, so a maintenance release of an older line no longer moves them. - A Windows profile keeps its policy identity. The MS-XCEP policy identifier and template OIDs belong to the profile, not to its CA: changing the policy or issuing CA no longer changes the identifier and OIDs registered clients cached (such a change made them fail with
0x8007000d), two profiles on one CA no longer share an identifier, and the default profile upgraded from 2.254 keeps the values 2.254 served. - FIPS mode allows what the loaded FIPS module approves. The
-fipsimage moves to the OpenSSL FIPS Provider 3.5.9 (CMVP certificate #5586), and ML-DSA, Ed25519 and Ed448 are allowed in FIPS mode when the module answers that it approves them, refused as before otherwise (RHEL 9, Ubuntu 24.04, 3.1.2), and the readiness check reportsmodule_known, false when it ran without a FIPS module; Ed25519 deployment keys and ML-DSA keys on an HSM or OpenBao stay refused, as they are signed outside the module. - FIPS mode: one RSA rule. An RSA key is approved when its length is even and from 2048 to 16384 bits (FIPS 186-5), at issuance, for HSM signatures, for EST server-side key transport and in the readiness check alike: a CSR with an RSA-8192 key is now signed in FIPS mode, and an odd length or one over 16384 bits is refused everywhere.
- FIPS image: hybrid post-quantum TLS. In FIPS mode the image's TLS offers X25519MLKEM768, SecP256r1MLKEM768 and SecP384r1MLKEM1024, the three hybrids the 3.5.9 module approves; X25519 and X448 alone stay refused.
- FIPS mode: blockers at startup. In FIPS mode each blocker of the readiness check is logged as an error with its remedy at every start and reported by the service status under
crypto.blockers, with the time of that check; UCM still starts and refuses the affected operations when used. - FIPS mode refuses HMAC secrets under 14 bytes. A webhook secret or TSIG key shorter than 112 bits is refused when saved (SP 800-131A), and a webhook secret, DNS provider secret, external EAB key or
SECRET_KEYthat short fails when used with the FIPS reason instead of a raw module error; the readiness check lists them, a shortSECRET_KEYas a blocker. - The database switch refuses a target this release cannot start. A target holding accounts from a release older than 2.254, or a 2.254 database without accounts, is refused (409) before
ucm.envis written, where 2.254 switched to it, and so is a SQLite file older than 2.254 when switching back. Retrying a switch cut short on the same target resumes it, where 2.254 asked for a reset (still required after a failed verification of the copy), and a target already started by this release that holds accounts gets its id counters moved up to the current database's. - WSTEP Kerberos requests without template are refused. A request on the Kerberos binding that names no template and whose EKUs match none of the profile was issued as a server certificate; it is now refused and audited (
windows_enrollment.refused), as AD CS does. UsernamePassword is unchanged, and so is the renewal of the requester's own valid certificate issued without template by Windows enrollment. - A WSTEP Kerberos renewal replaces the renewed certificate. The old certificate stayed valid beside its successor; it is now marked superseded as on the Certificate binding, a second renewal of it is refused, and after a change of the profile's CA a certificate of the former CA renews as such (audited
certificate.renewedwith the old serial and CA). - WSTEP names the directory's reason for a naked request. An autoenrollment request with an empty subject from an account without
userPrincipalName, or that the directory does not find, was refused withCSR subject must include a non-empty CN or a SubjectAltName; it now gets the directory's reason, in the fault and the audit. - WSTEP Kerberos renewals are audited for the requester's own certificate only. A request is audited
certificate.renewed, and emits thecert.renewedwebhook, only when the certificate that signs it is valid, of the same profile and template, and the requester's; signed by another certificate, or naming another template, it is not a renewal: an enrollment (certificate.issued), or a refusal when it names no template. - SCEP: a PKCSReq signed by a valid certificate this endpoint issued is now a renewal, as clients that renew with PKCSReq expect: the challenge is still compared but may have expired, and the old certificate is replaced; a CSR subject other than the certificate's is refused (see UPGRADE.md).
- EST re-enrolment and SCEP renewal refuse a certificate of a Windows enrollment profile and audit the refusal (
windows_enrollment.refused); it renews over WSTEP only. - Kerberos: each Windows enrollment profile holds its own keytab (
POST /api/v2/windows-profiles/{id}/keytab, entriesHTTP/<SPN host>@<realm of the profile>), and UCM composes the file the acceptor reads from the keytabs of every profile, disabled ones included. - XCEP and WSTEP Kerberos: a request is served by the enabled profile of its principal's realm (exact match), with that profile's CA, templates, connector and group, never by the default profile for another realm. A realm without a profile is refused with a SOAP fault and an audit entry.
- WSTEP: every certificate records the profile that issued it, and the Certificate binding renews it under that profile alone, only for certificates issued through Windows enrollment (see UPGRADE.md). Renewal with the TLS client certificate is limited to profiles whose issuing CA is the mTLS CA or under it.
- In-app help is now maintained in English, French, Russian and Chinese only; the other interface languages show the English help with a notice.
- Console: the sidebar gives way to the Atelier shell, a horizontal bar with a Modules menu (four groups, keyboard navigable), three pinnable modules kept in your preferences, a status bar fed by
/healthand a module picker on phones; the account menu now carries theme, language and sign out. - Console: the code, styles and tests of the pre-Atelier interface (old sidebar, grid dashboard, floating detail windows, light-theme legacy styles) are removed.
- Dashboard:
GET /api/v2/dashboard/activityanswers 500 on a database error instead of an empty list, so the activity card tells an outage from no activity. - Encryption: disabling it while
KEY_ENCRYPTION_KEYis set now succeeds and leaves the variable configured but unused, where it answered 409. - Configuration: the new
UCM_CONFIG_DIRvariable (default/etc/ucm) moves the directory from whichucm.envandmaster.keyare read and where the master key is created. - Backup: archives are written as format v4, a stream of independently authenticated chunks, so creating one, decrypting it and the checks before the restore plan no longer hold it whole in memory (the plan and the writes still hold every section it carries);
BACKUP_MAX_UPLOAD_MBnow bounds only archives written before 3.0 (v1 to v3), which are still restored, and the newBACKUP_MAX_ARCHIVE_MB(0, no limit, by default) can bound v4 archives. - Breaking: the SCEP CertRep is signed with the digest of the client's request (SHA-256, SHA-384 or SHA-512, never below the key's own), where it was always SHA-256; SCEP still serves only a software RSA CA; #397.
- Requests that name no key or hash follow one default on every issuance path: without a profile or policy it stays RSA-2048 with SHA-256, a profile or policy that refuses it moves to the first accepted type (RSA-3072, EC-P256, EC-P384) or answers 400 when none is. The hash follows the signing key, and a new intermediate CA never takes a hash below its parent's key or its own; #397.
- Templates: a fresh installation seeds the same system template list an upgrade completes (eight templates, Client Authentication and VPN Client among them) instead of an older seven-entry list.
- Breaking: a policy that requires approval now holds ACME, EST, SCEP and WSTEP requests in the approval queue, so a protocol request can stay pending until an approver decides; each client is told so in its protocol's terms (ACME
processing, EST 202 with Retry-After, SCEP PENDING, WSTEP pending RSTR and QueryTokenStatus), even on an auto-approving SCEP profile.
Fixed
- Health probes and a database that does not answer.
/health/liveread settings from the database through its request hooks and hung with it, so an orchestrator restarted healthy nodes; it now reads nothing from the database, and/health/readyanswers 503 within 5 seconds. - Start summary on PostgreSQL. The Docker image announced the unused SQLite path as its database on a single instance with
DATABASE_URL; it now names the engine and the database name. - OIDC attribute mapping to
groupsignored. An OIDC attribute mapping row togroupswas ignored: the mapped claim, by exact name or dotted path such asrealm_access.roles, is now the source of the groups in place ofrolesandgroups(see UPGRADE.md). The "Role resolution" log line no longer writes the whole role mapping and every group value. - A replacing restore leaves no ACME server state behind. The authorizations and challenges of an ACME account the archive does not hold were kept without an owner, the valid ones for good; they now go with the account, and the legacy CRL rows of a removed CA go with it.
- A replacing restore keeps the names of the archive. A webhook or ACME CA account that matched a row here under another name was restored under the name here; it now takes the archive's name.
- Bulk exports survive an unreadable row. The certificate and CA exports answered 500 in P7B and carried the damaged data in PEM, as the trust store did in PEM while its P7B dropped the entry silently and answered 400 when none was left; each now skips the row, logs it, lists its id in
X-Skipped-CertificatesorX-Skipped-CAs(first 100, with-Count) and answers 404 in both formats when none is readable. - AIA
.p7cbundle with an unreadable certificate. It answered 500; it now serves the readable certificates of the CA, logs the unreadable ones and answers 404 when none is readable. - Renewal during a CA record merge. An in-place renewal already stored answered with an error when the SKI/AKI repair merged the record of its CA right after; it now reports the renewal and regenerates the CRL of the record that remains.
- OIDC role mapping with Microsoft Entra ID. Role mapping rules now compare their keys with the
groupsandrolesread together from the ID token validated for the login and the UserInfo response, which Entra ID never fills, so a provider with rules no longer refuses every user; arolesclaim no longer hidesgroups. - SAML logins from Keycloak. An assertion that repeats an attribute name, as Keycloak's role list does with one attribute per role, was refused as invalid; their values are now merged in assertion order, role mapping reads them all and a repeated username or email attribute gives its first value.
- LDAP groups with a base DN at the domain root. With
memberoruniqueMember, groups were searched under the parent of the base DN, so a base DN such asdc=example,dc=comfound none and a user the rules named was refused; they are now searched under that parent when the directory holds it, as before, else under the base DN, and a search the directory does not answer with success is tried once more under the base DN, never read as "no group". - LDAP groups without a group filter. A provider with role mapping rules or required groups and an empty group filter read no group and refused everyone; the groups are now read (
memberOfneeds no filter,memberanduniqueMemberdefault to(objectClass=*)). - LDAP role mapping ignores case. A rule
pki-adminsdid not match the Active Directory groupPKI-Admins, which the required groups accepted; LDAP rules and groups now compare as caseIgnoreMatch does (case, edge spaces and inner space runs ignored), OIDC and SAML stay exact, and a rule key is saved without edge spaces (two keys the provider compares equal are refused). - A failed LDAP group read says so. When the groups could not be read (service account rights, timeout, referral, missing base) the login was refused as "no role mapping rule matches", or as not a member of the required groups; it is now refused with "your groups could not be read from the directory", in the response, the log and the audit, and like a missing required group it never locks the account by itself.
- The LDAP mapping test gives the login's verdict. It read the groups by its own code and showed "Resolved role: viewer" for a user the login refuses; it now reads them as the login does, shows the access line and the reason of a refusal (rule, required groups, unreadable groups, account disabled in the directory) with no role, and says what it does not check (
refusal,groups_readable,notein its response). - Active Directory disabled accounts in the mapping test. The
userAccountControlcheck never read the attribute's value, so a disabled account was only refused by the directory's bind. The mapping test now reports it disabled, and a login the directory binds despite that flag is refused. - Audit statistics on a SQLite database optimized early. After "Optimize database" had run on a nearly empty database, the period statistics of the audit log read the whole table; they keep the timestamp index now.
- Files after a restore. A restore no longer writes certificate, CSR or key files: once it has committed, the service rewrites from the database the files that are missing or differ and removes those of the rows the restore removed, and it refuses the restore with
409when it cannot write to these directories. - Backups of large installations. A backup held each section of the archive whole in memory and ran an RSA consistency check on every private key, which took most of its time; it now reads each table a page at a time, so its memory no longer grows with the number of certificates, and compares each key with its certificate by their public keys alone.
- Half an HTTPS pair in an archive. An archive whose HTTPS certificate or key file was missing at backup replaced only half of the pair served; the restore now keeps the HTTPS certificate in place and says so in its report.
- A restore never runs without its lock. On a filesystem where the backup lock cannot be taken, a restore was run anyway; it is now refused with
409. - Interrupted backups leave no temporary file. A backup whose process was killed left a
.ucm_backup_*.tmpin the backup directory; the next backup and each start remove those of backups no longer running. - No settings warnings at the first start on an empty database. Before the schema was built, the start logged
relation "system_config" does not existfor WSTEP, HTTPS and mTLS (no such tableon a SQLite file cut short); a database without that table now reads as one without settings, and mTLS keeps the mode of the previous start. - Packages: no bytecode of the previous version after an upgrade. The DEB and RPM post-installation remove the
__pycache__the service wrote under/opt/ucm, and the RPM keeps the dates of its sources: on Fedora 43 and EL10 every version carried the same file dates, so a source of unchanged size could run the previous version's bytecode. - SSH certificates: two workers or nodes signing under one SSH CA at the same time could issue the same serial number, so revoking one certificate in the KRL revoked the other too. Each signature now takes its own serial from the database.
- SSH certificates: an SSH CA whose counter trailed its certificates (after a restore, an import or a change of database) could issue a serial again; the database now holds each serial once per CA, every such path moves the counter past it, and certificates already sharing one are marked, shown and revoked together (see UPGRADE.md).
- Renewal: a renewal overtaken by another renewal of the same certificate, manual, bulk or scheduled, was refused as if the certificate were revoked. It is now refused as renewed by another request.
- Web interface: a request the browser cancelled could cut another download served at the same time, so the main script arrived truncated and the page stayed blank until reloaded. The server now empties the TLS error state before each read and write, so one dropped connection no longer breaks the others.
- Related certificates (RFC 9763): a member of a pair whose partner had just been renewed by another request was renewed alone and lost its link to that partner. The partner is now judged as the database holds it, so the pair renews together, and a partner revoked meanwhile leaves the member to renew alone.
- Related certificates (RFC 9763): a certificate given a newer partner by another request since it was read renewed with the previous one, or alone. Its own link is now read as the database holds it, and a link changed while the renewal runs answers 409 so the retry takes the right partner.
- Related certificates (RFC 9763): a certificate B unlinked by another request since it was read, its A having been renewed alone, renewed still naming A while linked to nothing. What the renewed certificate names is now decided on the row as the renewal locks it, and a B renewed naming nothing leaves its pair.
- Related certificates (RFC 9763): a certificate that names its own A and links back to its own B, renewed with that B, named the B's replaced certificate. It now names nothing its original did not.
- Related certificates (RFC 9763): revoking the newest B alone while its A was renewed alone could write the link of A back to an older B after the renewal had cleared it. The link is now handed over only to an A still locked and linked to the revoked B.
- Related certificates (RFC 9763): deleting a member while a newer B took its partner could clear or move the link that B had just set. The rows a deletion relinks are now locked and read again first.
- Automatic renewal: a renewal overtaken by another request (the certificate or its pair changed meanwhile) was counted as failed, audited as
certificate.auto_renewal_failedand mailed as a failure, though the next pass retries it. It is now counted as skipped. - Policies: editing a policy in the form dropped the rules the form does not show, such as
allowed_signature_schemesset through the API. They are now sent back unchanged. - HSM: syncing the keys of a provider whose secret this node cannot decrypt answered 500 without naming it. It now answers 400 and names the field.
- CA deduplication: merging duplicate records of one CA dropped the HSM key of a removed duplicate when the record kept had no key. The kept record now takes over that HSM key.
- Two-factor login: a TOTP secret under 112 bits or not base32 (restored or set outside UCM, never generated by it) made the code check fail with an error, recovery codes included. Its codes are now refused like wrong ones, recovery codes still sign in, and the log says the secret is unusable without showing it.
- French uses "AC" everywhere. The interface and the French help wrote "CA" and "la CA" for the autorité de certification; they now say "AC" with the right articles, and a test keeps the English acronym out.
- The help names the buttons the screen shows. The HSM, certificate and template guides, in all four help languages, cite Generate, Issue Certificate and Copy; the forced password change prompt is described as having Logout as its only other button. The unused
sso.addProviderlabel is removed from the ten locales. - A renewed session identifier no longer signs the console out. A request the page sent before a password change, a TOTP confirmation or the closing of the other sessions got 401 with the replaced identifier and sent the user to the sign-in page; the console now checks its current session first and sends a refused request again once. A request still being served when the identifier is renewed no longer writes back the record of the replaced identifier, which stays refused, nor sends that identifier or a cookie deletion to the browser.
- A restore signs out every session begun before it, its own included. With sessions on disk (on SQLite without Redis), the restore request and any request served during it wrote their session back once the restore had removed it, and with Redis no session was revoked at all; each restore now advances an epoch every session carries, so none begun before it is accepted afterwards, whatever the store and on every node.
- A start no longer loads every certificate into memory. While the HTTPS certificate is not a managed one, each start looked for it among the managed certificates by loading all of them with their keys; it now reads them in batches of 500, and the start of a 255,000 certificate database peaks at 250 MB instead of 1.8 GB (2.7 GB on PostgreSQL).
- A killed start no longer holds PostgreSQL locks. The migration session now asks the server to check its client every 10 seconds (PostgreSQL 14 and later) and uses the keepalives of the application's connections, so a start killed during a long migration releases its locks within seconds instead of when the server finishes the statement.
- Audit statistics on a large journal. On SQLite,
GET /api/v2/audit/statsread every entry of the journal to rank the users, whatever the period; it now reads the period only (about 1.3 s down to under 0.1 s with a million entries), and users with the same count are listed by name. - Faster audit chain verification.
GET /api/v2/audit/verifyreads the journal as plain rows instead of building an object per entry, which cuts its time by about a quarter on SQLite and by half on PostgreSQL with a million entries; the verdict is unchanged. - ACME proxy: the help and the proxy hints now point to the "Certificate Authorities (CA)" section by the exact title the screen shows, instead of "External CA Accounts".
- ACME, SCEP and CSR imports name the weak hash of a refused CSR. A CSR signed with SHA-1 or MD5 is refused with "CSR signed with an unsupported hash algorithm", as EST and WSTEP say, instead of a failed proof of possession, also on CSR upload and import, smart import and Microsoft CA signing (same HTTP codes); ACME keeps
badCSRand SCEP now answersbadAlg, and in FIPS mode SCEP names a refused SHA-1 CMS signature. RATE_LIMIT_ENABLED=1,yesoronno longer drops the route limits (sign-in, password reset and the default of 200 per minute and 2000 per hour). Onlytrueturned on the route limits while the tiered limiter also took1,yesandon; both now read the variable the same way (any case), and an empty value keeps the default, on.- WebSocket events through Redis on a Unix socket. With a
unix://REDIS_URL, the event relay passed TCP keepalive options that a Unix socket connection refuses, so every WebSocket connection failed and no browser received real-time events; they now go through Redis as with a TCP address. - A WSTEP renewal is served beside an imported copy. When the same certificate was also stored under a second record of its CA and that copy was read first, a Kerberos renewal without template, a renewal authenticated by the certificate (TLS or WS-Security) and its approval were refused; the row that carries the Windows profile now decides.
- Two SQLite processes creating the default Windows profile at once. The one that lost the write lock logged "Kerberos keytab not restored" at startup; it now waits for the other's profile and uses it.
- A first start cut short completes its schema at the next one. On SQLite, a fresh install stopped while its tables were being created kept, at every later start, the tables and indexes it had not reached; the start now creates whatever the model holds that the database lacks.
- The LDAP mapping test reads group names as the login does. A group or base DN with an escaped comma (
CN=Ops\, Paris,...) gave the test screen a truncated group, so it predicted a role or a refusal the real login would not give; it now honours the escapes. - A certificate whose extensions cannot be read is refused at import. A certificate that repeats an extension or carries a malformed one was stored as a CA, or answered 500; every import path (file, paste, smart import, OPNsense, CA certificate install, cross-certificate attach, restore-primary, trust store, converter, mTLS certificate import) now refuses it with a 400 or a per-item error that names the extension.
- Form labels name their field. Text fields, text areas and selects are now tied to their label, so assistive technologies announce the field and a click on the label focuses it.
- The page language follows the interface language. The
langattribute of the page changes with the language chosen, so screen readers and browsers read it in that language. - Recovery codes are shown when TOTP is enabled. The enrolment dialog closed on confirmation and the codes the server returned were never shown; the dialog now lists them until they are marked as saved.
- A new CA appears unfolded in the hierarchy. An intermediate created under a collapsed root was hidden until the root was opened by hand; every newly seen CA is now expanded.
- FIPS on Ubuntu 24.04: the package and the release notes now tell to enable Ubuntu Pro's
fipsstream, which ships the binaries certified under CMVP #5115, instead offips-updates, whose patched versions are not the certified binary. - Release notes: a release lists its own section, or
[Unreleased]for a release candidate without one, followed by a link to the CHANGELOG at the tag instead of earlier sections; a stable release whose section is missing, or a release candidate with neither, now stops instead of publishing notes from older versions. Release tooling accepts three-segment maintenance versions (2.254.2) and writes the npm form of the version in both npm files. - Automatic update install: a release whose package was published before its
.sha256file got a failedsettings_updateaudit entry and was never installed automatically; it now waits, audited once per version, and the next check after the file appears installs it. - Scheduled CRL: whether a CRL is due is judged on the CRL served, the one with the highest number, instead of the one stored last.
- Overview links open their object. The links of the work queue and the bell now open the approval, SCEP or key recovery request, ACME order or CA they name (new reads
GET /scep/requests/{id},/key-recovery/{id}and/acme/orders/{id}), and say so when it was handled meanwhile. - Overview states in the user's language. Request states, CA statuses and the last run of a protocol are no longer shown in English, and the scheduler line shows failed tasks and an unpublished state instead of the latest run only.
- A switch no longer takes a copy that failed its verification. A target holding accounts copied by an earlier switch whose verification failed or was cut short is emptied and copied again from the current database, then verified, and one left by a data migration that did not pass its verification is refused until it is reset; before, the second attempt took its accounts as they were.
- A failed switch says whether retrying helps. A switch stopped by a lost connection, a deadlock or a locked database answers 503 and says that retrying resumes it, or starts it again when nothing was written yet; any other failure of the copy answers 500 and asks for its cause to be corrected first, where both promised a resume as they were.
- A CA certificate from another issuer no longer keeps the old parent. Replacing or renewing a CA certificate links the CA to the stored CA whose name and key both match the issuer of the new certificate, else keeps the old parent only while neither the issuer name nor the key the certificate names disproves it (a dropped link is audited as
ca_parent_unlinked), the rule the hourly chain repair applies to links left by an earlier version; a revoked CA keeps its parent, and an installed renewal no longer refuses a parent whose name differs only by case. - CA hierarchy: a stored CA certificate that repeats an extension no longer stops the hourly chain repair, nor fails the display or re-import of the CAs under it; their links to it are left as they are.
- A comma in a common name is kept. A CN such as
Doe, Johnwas cut toDoe\in the certificate and CA lists and details (subject, issuer and the other name fields), the smart import, the search results, the enrolled user certificates, the system trust store sync and discovery events; it is now read whole, and a CA so named still counts the certificates that name it as issuer. - Revocation: while lifting a hold on a CA with delta CRLs, OCSP answered
revokedwith reasonunspecifiedand a complete CRL listed the certificate without a reason. Until the hold is lifted, both now answercertificateHoldwith the date of the hold, a final revocation recorded for the serial still prevails, and only the delta CRL saysremoveFromCRL. - Docker: stopping the container now stops UCM cleanly. The server never received the stop signal and was killed when the grace period ran out, so a node in HA kept its row until its heartbeat expired.
- FIPS mode: a CA on an HSM key other than RSA 2048+ or ECDSA P-256/P-384/P-521 (RSA under 2048 bits, Brainpool, secp256k1) kept signing after the switch; its certificates and CRLs are now refused, OCSP answers
unauthorized, andPOST /api/v2/hsm/keys/<id>/signanswers the 400 FIPS problem instead of 500; creating a CA on such an existing key gets the same refusal. - INWX DNS provider: a two-factor shared secret whose length is a multiple of 8 characters, such as the usual 32, sent a wrong code and the login failed; it is now decoded as base32.
- FIPS readiness check: it lists every HSM signing key FIPS mode refuses to sign with (Brainpool, secp256k1, RSA under 2048 bits, unknown types), judged as the signature is, on the public key UCM holds for the key or else on its type.
- FIPS mode: a CA or certificate imported with an ML-DSA key was accepted where generating one was refused; it now gets the same FIPS refusal unless the loaded module approves ML-DSA (upload, CA import, smart import).
- Restore: a request or task that committed on the same node while a restore ran had its commit silently turned into a flush and lost; only the restore's own session is now held to one transaction.
- Settings:
max_login_attemptsunder 3, the settings screen's minimum, is refused (400) instead of locking every login at its first failure, and a lower value already stored is applied as 3. - CA import: a CA whose RSA key is under 2048 bits is refused when it arrives with its private key (upload, PKCS#12, the key of a pending request, adding a key to an existing CA with
POST /api/v2/cas/<id>/key, smart import, OPNsense import), with a message giving the size and the minimum, and so is a CA key that cannot be read; the same CA without a key is still imported. In a batch only that CA is skipped and named. - OPNsense import and smart import chain a CA under the stored CA whose key signed it, whatever the order of the batch, instead of leaving it unlinked or taking a CA of the same name. A CA whose parent signature cannot be verified here (a parent key type the server does not handle, such as SLH-DSA or GOST) is imported without a parent, where smart import used to attach it by name.
- CA import: a CA or certificate whose issuer is stored both as a root and as a cross or link certificate of the same key is now chained under the self-signed root, on SQLite and PostgreSQL alike, including a certificate installed on a pending CA and certificates linked by Smart Import, OPNsense import and the hourly chain repair. The chain status of a CA with no recorded parent and the chain added to an export follow the same rule.
- Behind a trusted reverse proxy, the pending Kerberos handshakes of XCEP and WSTEP are keyed by the client address instead of the proxy's, so two clients no longer share one. The keytab composed from the Windows profiles is refused above 1 MiB with a readable message.
- Kerberos: a Kerberos Microsoft CA connection pointed
KRB5_KTNAMEat its own keytab for the whole process, so XCEP and WSTEP could accept with the wrong keys; the acceptor now names its keytab itself. A ticket for a realm without key is logged as a warning instead of an unrelated GSSAPI error. - SCEP with Intune: the Entra token and the discovered service endpoint are kept between enrolments (one client per app registration), and a 401 or 403 from Graph or the validation service makes the next request ask for a new token.
- SCEP: a
PKIOperationsent by GET with an unencoded+in itsmessageis read correctly instead of failing. - Expiry webhooks: after an outage,
certificate.expiredis announced for what expired since the last pass, at most 30 days back, on a single node as in HA (a single node used to cover the last 24 hours only); the log gives the number of older expiries left aside. - Certificates: lifting a hold now emits
certificate.unheld(webhooks, WebSocket) and lifts the hold of the other member of a related pair too, unlessunhold_relatedisfalse, reporting it asrelated_unheld; a body that is not a JSON object is refused (400 or 415), and a pair on one CA gets a single delta CRL. - Import: an imported CA now announces
ca.createdandca.updated, and a submitted CSR carries the acting user incsr.submitted. - Events: uploaded and imported CSRs emit
csr.submitted, a CA signed from a CSR emitsca.created, an OCSP responder or EST key generator renewal emits its issued or renewed event, and approving a pending Microsoft CA request announces the certificate it imports. - Notification bell: a failed restore shows again, as a failed backup does.
- ACME: a certificate obtained over ACME is also revoked at its CA when it is revoked as a user certificate, with its mTLS enrolment or user, as the related member of a pair or by its key at UCM's own ACME server. A revocation is announced as sent again only once that retry is recorded, a database error no longer leaks its SQL into the answer, and a bulk or approved hold no longer warns to revoke at the CA.
- WebSocket: without Redis, certificate and CA events born in the scheduler process, such as automatic renewals, now reach the browsers, relayed to the web workers through the database.
- Import: replacing a certificate through
POST /api/v2/certificates/importwith a new serial now fires the renewal event, so its deploy bindings, webhooks and HTTPS binding follow as after an internal renewal (discussion #440); re-importing the same serial stays silent. - Approvals: a certificate issued by approving a request now fires
certificate.issuedlike the issue form, so webhooks, notifications and deploy bindings see it. - AD CS: a certificate AD CS signs for a CSR or a renewal fires
certificate.issuedorcertificate.renewed; approving a pending request from UCM writes it into the request's own row, where a later status check used to add a second row. - ACME client and proxy: a certificate obtained upstream fires
certificate.issued, orcertificate.renewedfor a renewal stored in a new row; the proxy no longer stores a second row when a client downloads the same certificate again. - Import: importing a new certificate, or a certificate that completes a pending CSR, fires
certificate.importedfrom the import page as the Smart Import does, and so do the mTLS certificate import and the OPNsense import for each certificate they create. - Renewal: when EST, SCEP or WSTEP renews a certificate into a new row, its deploy bindings move to the new row in the same commit and the renewal is deployed, except a binding that pushes the private key and the HTTPS listener binding, which stay on the old row (UCM holds no key for the new one) with a logged warning. An approved SCEP renewal fires
certificate.renewedinstead ofcertificate.issued. - AD CS: when the status check of a pending request imports a certificate that replaces an existing one, its deploy bindings, webhooks and HTTPS binding now follow as after a renewal.
- Syslog: a change or a deactivation of the remote syslog target now reaches the scheduled tasks within seconds; their audit entries kept going to the previous target until the service restarted.
- Backup: a restore now signs a fresh CRL (and delta CRL) for every CA that publishes one and whose key matches its certificate, instead of serving the previous one until its renewal threshold. A CA that cannot sign (HSM missing, mismatched key) or whose delta failed is named in the restore result.
- CRL: no CRL is signed any more with a key that is not the CA certificate's, whether by the scheduler or a manual full or delta regeneration, which answers 409 and says why.
- Settings: changing the HTTP protocol port now asks for the service restart it needs (a rolling restart in HA), and when
HTTP_PROTOCOL_PORTin the service environment pins the port, as the DEB and RPM packages do, the field is read-only, shows the port really bound and a save says the stored value is not used. - Issuance: a certificate created with
POST /api/v2/certificates(the certificate form), issued from the form once its approval was granted, enrolled through SCEP or renewed in place, a cross-certificate, and the certificates UCM issues to its own TSA signer, EST key generator and renewed OCSP responder were never counted in their CA's issued certificates; they now are. - Renewal: an EST, WSTEP or SCEP renewal that waited behind an in-place renewal of the same certificate archived the certificate just renewed; it is now refused, and the OCSP responder renewal task leaves such a certificate as it is; an in-place renewal of a certificate archived meanwhile answers 409, an issuance linked (RFC 9763) to a certificate renewed meanwhile answers 400.
- Issue form: the preselected key now comes from the server default for the chosen CA and template (new
default_key_typeinGET /api/v2/policies/effective) instead of a fixed RSA-2048, and a key you picked yourself is kept. - Discovery: certificate validity dates are stored as UTC whatever offset the probe reports, and the expired and expiring counts compare in UTC, so PostgreSQL with a non-UTC session time zone no longer shifts them.
- Concurrency: two workers or nodes issuing under one CA no longer lose counts of the CA's issued certificates, and the database refuses a second certificate stored with the same serial under one CA (the upgrade lists existing duplicates and reports them in
/healthinstead of creating the index). A discovery scan, a delegated OCSP responder renewal and an ACME proxy challenge now behave the same whichever worker or node handles them. - Certificates: a certificate could be stored twice under one CA, with a decimal and a hexadecimal serial, past the unique serial index; every path now stores the decimal serial, the upgrade converts existing rows (rows of one CA that would then share a serial are listed, not converted) and search finds a serial typed in hex.
- Backup: a manual backup or a restore that failed left nothing in the audit log; it now records
system_backuporsystem_restorewithsuccessfalse and a reason, as the scheduled backup does. - SSO: Link to SSO refused every account already linked or left without provider, the remedy sign-in refusals named, and Unlink then Link made an account without password set one at its next SSO session. Attaching a provider now works on any account, keeps a password it has, asks none of an account without, closes the account's sessions, can remove its local password, and its audit entry names the providers the account had.
- SSO: a sign-in whose email belonged to an account of another provider or left without provider by another or an unknown IdP (after a restore, for instance), or to an account an administrator linked while the OIDC IdP marked that email unverified, received a second account with the default role. It is now refused (
sso_email_conflict) with an audit entry naming the account. - SSO: a sign-in refused over a name gave the same reason for every account; the log and audit entry now say whether the account is local, linked to another provider, bound to another identity or without provider, and what repairs it. A provider with no rule and no default role now says so on the login page (
sso_no_default_role). - Users: a CSV import refused a local account whose e-mail an SSO account used, which a single creation accepts, and answered
500when another request took one of its e-mails or usernames first; it now accepts that e-mail and answers409naming the e-mail or username taken. - SSO: a second attribute mapping row with the same attribute name silently replaced the first one on save; the row is now marked and the provider is not saved until one of them is changed or removed.
- SSH: the KRL of a CA failed to generate once a certificate of serial 0 (an imported one) was revoked; such a certificate is now listed as OpenSSH lists it, by its key ID under that CA, and the CA's other certificates with that key ID are revoked with it.
Security
- Trust store: a file imported with its private key (a
server.pem) was stored whole, and the key came back to every account allowed to read the trust store, in the PEM bundle and in backups. An entry holding a private key is now refused, entries keep only their certificates, the upgrade and the restore remove a key already stored (audited, see UPGRADE.md), and no export hands out anything but certificate blocks. - Account: the password reset by e-mail set a local password on an account created by an SSO provider, which then opened without the provider. Such an account gets no reset link any more, and a reset token sets no password on it.
- Account: passwords sent at the same time to the password or LDAP login, on one node or several, were all checked before the lockout, and a success could leave another node's lock in place. Each attempt is now counted before its password is checked, at LDAP on the account the directory entry opens whatever name was typed, so at most
max_login_attemptsare checked before the account locks, alerted once, and a success clears the stored count. - Account: a login by client certificate or passkey reset the password's failed attempt count, giving whoever guesses the password a fresh set of tries. The count and the lock now belong to the password and the LDAP bind alone.
- Account: an LDAP sign-in whose password is right but which opens no session (account disabled, not in the required groups, no role mapped) now counts toward
max_login_attemptsalongside wrong passwords, though such sign-ins alone never lock the account. A password set by an administrator or changed in a session now clears the count and the lock, as a reset by link does. - Integrations: a stored secret the database key cannot decrypt (saved under another key) now stops its integration with an error naming it: a webhook was sent unsigned or without its auth token, a mail without SMTP login, an SSO token request, a Microsoft CA or an HSM provider connection without its secret (a Google Cloud KMS provider fell back to the node's default credentials).
- Integrations: a saved secret this node cannot decrypt, sent back masked or blank, now holds its target like a readable one, so moving an HSM provider, AD connector, SSO provider, Microsoft CA or SMTP server without typing it again is refused, and a DNS provider's credentials change only when entered again in full, every required field, and a secret when the provider has one, in one request (an optional secret left out is then dropped). The AD connector's test, health probe and domain SID read, the Intune app tests, the approval of an Intune SCEP request and a DNS provider's challenges, cleanups and tests (except a manual provider, which reads none) name such a secret instead of reporting it missing or a generic failure; an enabled AD connector holding one saves only with its bind password typed again or switched off.
- FIPS image: in FIPS mode the container refuses to start when its FIPS module differs from the one the image was built with, when the image's integrity value for it is missing or malformed, or when
OPENSSL_MODULESis not the absolute path of that module's directory; 2.x before 2.254.3 ranopenssl fipsinstallon whatever module it found and so accepted a replaced one (see UPGRADE.md). The checked directory then holds for the whole process: anOPENSSL_MODULESinucm.envdoes not move it. - Account: a password checked again in an open session (password change,
PUT /api/v2/users/{id}andPOST /api/v2/users/{id}/reset-passwordon one's own account, turning off the security key as second factor) now counts its failures per account, apart from the login's; pastmax_login_attemptsit answers 429 withRetry-After, the right password included, untillockout_durationhas passed. - Account: confirming or disabling TOTP, regenerating recovery codes, adding, removing, enabling or disabling a security key (an administrator's change included) and turning the security key as second factor on or off now close the account's other sessions, as a password change does; the audit entry gives how many. Under enforced 2FA without TOTP, turning off the security key as second factor is refused (403,
last_second_factor), as removing the last key is. - Login: failed second-factor (TOTP, recovery code, security key) and passkey attempts are counted in the database, so the lockout applies on every node, survives a restart, and parallel attempts never get more tries than
max_login_attempts. An attempt counts when it is made rather than when it fails, and a count left idle forlockout_durationis forgotten. - Backup: a merge restore no longer lowers a revocation made here, final or a hold, of a certificate or a CA, nor turns its revocation record into a hold; only a graver revocation in the archive replaces it.
- Revocation: the upgrade deletes once the suspension records that lifting a hold used to leave behind, unless a 2.x upgrade already did.
- Revocation: a hold its CA recorded but the certificate does not show yet can be lifted, also as the related partner of a lifted certificate, and the certificate details and list show the revocation the CA recorded.
- Users: deleting a user, alone or in bulk, left the certificates they had enrolled for mTLS valid on the CRL and at the OCSP responder; they are now revoked (cessationOfOperation) and deleted as when an enrolment is removed; a certificate that cannot be revoked does not block the deletion but is named in the answer (
unrevoked_certificates, orunrevokedper user in bulk), the warning and the audit log. - Users: on SQLite, a new account took the id of a deleted account when it held the highest id, with the rows that still named it; ids are no longer reused, a backend switch included.
- Dashboard:
GET /api/v2/dashboard/system-statusreturned the mail (with its host), webhook and auto-renewal status to any authenticated caller; these entries now needread:settings, like the same witnesses on/dashboard/health. - ACME client: revoking a certificate obtained from an ACME CA (single, bulk or approved request) only revoked it in UCM while the CA still answered "good"; the CA is now told (RFC 8555 §7.6) with reasons 1, 3, 4, 5 and 9 (privilegeWithdrawn included) and retried by the scheduler until it confirms, and the answer says when the revocation stayed local (
acme_local_only), a hold included.
Upgrade notes
- FIPS image: the module becomes the OpenSSL FIPS Provider 3.5.9 (certificate #5586) without any step; TLS 1.2 peers without Extended Master Secret are refused in both directions (see UPGRADE.md).
- CA import: a CA with an RSA key under 2048 bits is no longer accepted together with its private key; import it without the key (read-only, for chains) or replace it.
- Kerberos: a keytab merged for several realms is refused on upload; upload each realm's keytab on its own profile.
- First access: scripts that logged in as
admin/changeme123after a fresh install must provide the password or read the generated file; the DEB and RPM upgrade removes theINITIAL_ADMIN_PASSWORD=changeme123line 2.x wrote, and an existing administrator is left as it is. - Discovery: on PostgreSQL with a server time zone other than UTC, discovery dates stored before this release are in local time; run a scan of the discovery targets after upgrading to rewrite them.
- Windows enrollment: the XCEP, WSTEP and Kerberos settings become the profile
defaultwith an explicit realm (connector base DN, else SPN, else a single-realm keytab), migrated with the same policy ID and template OIDs; without a realm to take, Kerberos alone is refused until one is set, see UPGRADE.md. - Windows enrollment: a Kerberos request naming no template whose EKUs match none of the profile is refused instead of issued as a server certificate; name the template or ask for its EKUs.
- HSM: creating an X.509 CA on an HSM key (
hsm_key_id, orhsm_provider_idfor a new key) now requiresread:hsmbesideswrite:cas, as for an SSH CA; add it to custom roles and API keys that do this (see UPGRADE.md). - HSM: a 2.254 database where two X.509 CAs share an HSM key is refused at upgrade, nothing written, naming the CAs and the key; keep one CA per key in 2.254 first (see UPGRADE.md).
- Dashboard: an API key without
read:settingsno longer receivessmtp,webhooksandauto_renewalfromGET /api/v2/dashboard/system-status; add the scope to keys that read them (see UPGRADE.md). - Certificates:
serial_numberof certificates issued from the form, renewed in place, issued on policy approval, issued by a Microsoft CA or used as TSA signer is now returned in decimal like the others; the UI still shows colon hex. - SSO: a sign-in whose email belongs to an account of another provider is refused instead of receiving a second account; if it is the same person, attach this provider to the existing account, otherwise create an account under the name the provider sends and attach the provider to it, removing its local password.
Earlier releases: CHANGELOG.md
Installation
Docker (Recommended)
# From Docker Hub
docker pull neyslim/ultimate-ca-manager:3.0-rc1
# Or from GitHub Container Registry
docker pull ghcr.io/neyslim/ultimate-ca-manager:3.0-rc1
# Run
docker run -d -p 8443:8443 \
-e SECRET_KEY=$(openssl rand -hex 32) \
--name ucm neyslim/ultimate-ca-manager:3.0-rc1Debian/Ubuntu
wget https://github.com/NeySlim/ultimate-ca-manager/releases/download/v3.0-rc1/ucm_3.0.rc1_all.deb
sudo dpkg -i ucm_3.0.rc1_all.deb
sudo apt-get install -fFedora
wget https://github.com/NeySlim/ultimate-ca-manager/releases/download/v3.0-rc1/ucm-3.0.rc1-1.fc43.noarch.rpm
sudo dnf install ./ucm-3.0.rc1-1.fc43.noarch.rpmRHEL, AlmaLinux, Rocky Linux 9
wget https://github.com/NeySlim/ultimate-ca-manager/releases/download/v3.0-rc1/ucm-el9-3.0.rc1-1.el9.noarch.rpm
sudo dnf install ./ucm-el9-3.0.rc1-1.el9.noarch.rpmRHEL, AlmaLinux, Rocky Linux 10
wget https://github.com/NeySlim/ultimate-ca-manager/releases/download/v3.0-rc1/ucm-el10-3.0.rc1-1.el10.noarch.rpm
sudo dnf install ./ucm-el10-3.0.rc1-1.el10.noarch.rpmFIPS mode
The -fips image runs on the OpenSSL FIPS Provider 3.5.9 (CMVP certificate #5586), which approves ML-DSA, Ed25519 and Ed448; the packages run on the FIPS module of RHEL 9 (#5573) or Ubuntu 24.04 (#5115), which do not. Full guide: FIPS Mode.
Docker, any Linux host (amd64, arm64):
docker run -d -p 8443:8443 -e UCM_FIPS=1 \
-v ucm-data:/opt/ucm/data -v ucm-config:/etc/ucm \
--name ucm neyslim/ultimate-ca-manager:3.0-rc1-fipsUbuntu 24.04 with Ubuntu Pro (sudo pro enable fips, then reboot). The binary certified under #5115 is the one of Ubuntu Pro's fips stream; fips-updates carries later, patched versions that are not the certified binary:
sudo apt-get install -y gcc pkg-config python3-dev python3-venv libpq-dev libxml2-dev libxslt1-dev \
zlib1g-dev libssl-dev libxmlsec1-dev libxmlsec1t64-openssl rustc-1.91 cargo-1.91
sudo install -d -m 750 /etc/ucm && echo 'UCM_FIPS_BUILD=1' | sudo tee /etc/ucm/ucm.env
sudo apt install ./ucm_3.0.rc1_all.debRHEL, AlmaLinux, Rocky Linux 9 (sudo fips-mode-setup --enable, then reboot; on RHEL, enable CodeReady Builder instead of CRB):
sudo dnf config-manager --set-enabled crb
sudo dnf install -y python3.12-devel gcc pkgconf-pkg-config libpq-devel libxml2-devel libxslt-devel \
openssl-devel xmlsec1-devel xmlsec1-openssl libtool-ltdl-devel rust cargo
sudo install -d -m 750 /etc/ucm && echo 'UCM_FIPS_BUILD=1' | sudo tee /etc/ucm/ucm.env
sudo dnf install ./ucm-el9-3.0.rc1-1.el9.noarch.rpmThe package builds its crypto libraries against the system OpenSSL (about five minutes), sets UCM_FIPS=1 on a host in FIPS mode and prints a readiness report. Fedora, Debian and other distributions are not supported in FIPS mode: use the Docker image.
Silent/Automated Install
# Skip firewall prompts for CI/automation
sudo UCM_PORT=8443 UCM_FIREWALL=no dpkg -i ucm_3.0.rc1_all.debFirst Access
- Username:
admin - Password: none by default. Set
INITIAL_ADMIN_PASSWORD_FILE(orINITIAL_ADMIN_PASSWORD) before the first start, or read the generated one withsudo cat /opt/ucm/data/initial-admin-password(Docker:docker exec <container> cat /opt/ucm/data/initial-admin-password); a generated password must be changed at first login.