writeDomainsToCompose returns a shell fragment that is executed as part of the
compose build script. On error it interpolated error.message (which embeds the
user-controlled serviceName/host) directly into an echo, so a serviceName like
$(cmd) executed as root. Escape the message with quote().
The /api/providers/github/setup callback performed privileged writes (persisting
App client_secret/webhook_secret/private_key, re-pointing installations) with no
session check, deriving the write target (organizationId/userId on gh_init,
githubId on gh_setup) from the attacker-controlled state parameter. An
unauthenticated request, or a member with no gitProviders permission, could plant
or re-point a provider in any organization.
Require a valid session, gate on the gitProviders permission (same guard the tRPC
github router uses), derive the create target from the session, and verify the
gh_setup provider belongs to the caller's organization.
Host-level schedules (server / dokploy-server) run their script as root on the
host. The owner/admin gate only ran in the no-service branch, so a member could
attach an accessible applicationId to a dokploy-server schedule and skip it,
gaining root via schedule.runManually.
Extract assertHostScheduleAccess into the schedule service and call it before
the service-access branch in create/update/delete/runManually so the host-level
authorization always applies.
- application general 'Open Terminal' now passes serviceId (applicationId).
- The docker/terminal DockerTerminalModal now forwards serviceId to its Terminal.
- compose container rows thread serviceId to their terminal + logs modals.
This lets a member with service access (but without the docker permission) open
the terminal / read logs of their own service's containers from every service
view, matching the backend service-first authorization.
When a container is opened from a service page (serviceId present), gate on
service access alone — matching application.readLogs — instead of requiring the
docker permission first. This unblocks a member who has the service but not the
canAccessToDocker permission (or the server it runs on) from reading its logs /
opening its terminal. Generic Docker-overview ops (no serviceId) still require
docker permission + server access.
- Reorder canAccessDockerOverWss: docker permission -> service access (if
serviceId) -> server access, so the service grant is the primary signal.
- Pass serviceId from the service log views (application + postgres/mysql/mariadb/
mongo/redis/libsql + compose) so container logs opened from a service page are
gated by checkServiceAccess, matching the terminal path.
Completes GHSA-qf9j service-level dimension: when a container terminal/logs/stats
is opened from a service page, the frontend now passes serviceId, and the wss
authorizer calls checkServiceAccess so a member restricted to specific services
cannot reach another service's containers. Generic Docker-overview opens have no
serviceId and keep the docker-permission + server-access gate.
Verified against the local WebSocket: a member (docker:read=false) is rejected
with close code 4003; owner is allowed.
Closes GHSA-qf9j-c9p4-r4xp
The /terminal, /docker-container-terminal, /docker-container-logs and
/docker-stats WebSocket handlers only checked session + organization, so any
authenticated member could open a root shell / read logs of any container, and
the local (serverId=local) branch reached a root SSH shell on the control-plane
host.
Adds an authorization layer (server/wss/authorize.ts):
- container ops (terminal/logs/stats): require the docker permission (owner/admin
or a member granted canAccessToDocker); for a remote server, require the server
be accessible to the caller.
- host/server terminal: the local host root terminal is restricted to owner/admin;
a remote server terminal is gated on server access.
Closes GHSA-c68r-7wg9-p7v2, GHSA-899j-cjwp-v4gw, GHSA-7r6p-v9gw-pwc8, GHSA-qf9j-c9p4-r4xp
- cluster.removeWorker: input.nodeId (z.string(), no regex) was interpolated raw
into 'docker node update/rm ${nodeId}' — now shell-quoted.
- swarm image upload (getRegistryCommands): registryTag / imageName were
interpolated raw into 'docker tag'/'docker push' and an echo. registryTag is
built from username and imagePrefix, which have no schema regex, so it was
injectable — now shell-quoted. (The docker login already used
safeDockerLoginCommand, so credentials were already safe.)
- gpu-setup: nodeId (derived from 'docker info', not user input) escaped as
defense-in-depth.
Closes GHSA-4mfc-grxw-6858, GHSA-hfwh-69ch-gv47, GHSA-prwq-2mcm-mvhr
- composePath / appName are now passed through shell-quote in createCommand,
getCreateEnvFileCommand and services/compose.ts deploy commands, instead of
being interpolated raw into 'docker compose'/'docker stack' shell commands.
- compose.command: sanitizeCommand was cosmetic (trim + strip quotes). It now
rejects shell control characters (; & | ` $ () {} <> newline), which a normal
docker compose CLI line never contains, blocking breakout into host commands.
Closes GHSA-8r5w-vqjr-8c44, GHSA-5xv2-7f8w-9j5c, GHSA-qh6h-669j-77rw
The backup and restore command builders interpolated database name / user /
password directly into a 'docker exec ... {bash,sh} -c "..."' string executed
via execAsync / execAsyncRemote. Because the values sit inside the outer shell's
double quotes, a simple quote() is insufficient: the outer shell expands $(),
backticks and $VAR before the inner quoting applies (double-nested shell).
Values are now passed to the container as environment variables (docker exec -e
VAR=<shell-quoted>) and referenced as "$VAR" inside a single-quoted inner
script, so they never enter the inner command text and cannot break out of
either shell layer. The rest of each command (pg_dump/mysqldump/etc., flags,
| gzip) is unchanged.
Closes GHSA-qc73-mp78-4833, GHSA-qf8x-98cv-92qh, GHSA-ww4j-wjrr-rq8v, GHSA-7m3w-rm5f-h4fr, GHSA-f7mp-9jfp-mjrr
The deploy functions for postgres/mysql/mariadb/mongo/redis/libsql interpolated
the user-settable dockerImage field unquoted into 'docker pull ${dockerImage}'
executed via execAsyncRemote (SSH) on the remote server path. Now passed through
shell-quote. The local path already used pullImage() (execFile-based) and is
unaffected.
Closes GHSA-6jrh-8qmg-jj3p
- dockerImage (buildRemoteDocker) -> docker pull / echo
- dockerContextPath (docker-file builder) -> cd
- publishDirectory (nixpacks builder) -> docker cp source/dest paths
These fields were interpolated unescaped into shell commands run via execAsync /
execAsyncRemote during deployment. All are now passed through shell-quote.
Registry credentials already went through safeDockerLoginCommand (unchanged).
Closes GHSA-g9cg-4mmj-mh7p, GHSA-jxxj-gmpx-h5rj, GHSA-qjrc-g63x-qhp9, GHSA-98j8-6vjr-c3xw
swarm.getNodes/getNodeInfo/getNodeApps/getAppInfos accepted a caller-supplied
serverId and ran remote Docker Swarm reads against it with no organization
check, exposing another tenant's swarm topology cross-org (getContainerStats
already had the check; the others did not). getNodeInfo additionally
interpolated nodeId unescaped into 'docker node inspect ${nodeId}'.
Adds an org-ownership assertion to the four unscoped handlers and passes nodeId
through shell-quote in getNodeInfo.
Closes GHSA-jj6h-388v-9rwm
findServerById eagerly loads the sshKey relation (needed for server-side SSH
operations) including the plaintext privateKey. server.one and server.remove
returned that record to the client, exposing the private key to any member
with server:read regardless of canAccessToSSHKeys.
Adds redactServerSshKey() in the server service and applies it to the server.one
and server.remove responses. No client feature consumes the private key (the SSH
key management UI uses the dedicated sshKey router), and server-side callers keep
using findServerById directly, so behaviour is unchanged.
Closes GHSA-w9cp-jqfw-4xj9
The getBranches inputs type the provider id as optional, so only fetch and
authorize when an id is actually present; the underlying getBranches already
returns [] when no id is supplied.
The .one / getRepositories / getBranches / testConnection handlers for
github, gitlab, gitea and bitbucket resolved a provider by bare id under
protectedProcedure and returned the full row (OAuth tokens, app private
keys, webhook secrets) with no organization or entitlement check, letting
any authenticated user read another org's git-provider credentials (IDOR).
Adds assertGitProviderAccess() in the git-provider service (rejects cross-org
with NOT_FOUND and non-entitled same-org with FORBIDDEN, reusing the existing
getAccessibleGitProviderIds entitlement logic) and calls it in every read
handler before returning secret-bearing data.
Closes GHSA-66r5-5m5f-57vg
Updated the PopoverContent components across various provider files to ensure consistent width and padding. Enhanced the display of repository and branch names by adding truncation to prevent overflow in the UI. This change applies to Bitbucket, Gitea, GitHub, and GitLab provider components, as well as their respective compose components.
The repository CommandItem used the repo name as its cmdk value and
the check icon compared only names, so two repos with the same name in
different owners/orgs showed the selected checkmark and hover on both
entries. Key the item by owner/name and include the owner in the
selected comparison. GitLab already keyed by URL and is unaffected.
Fixes#4793
The Server Domain form rejected an empty value ('Invalid domain name'),
making domain assignment a one-way operation. The backend already
removes the Traefik router and clears the host when it receives an
empty host, so only the client-side validation blocked removal.
Allow an empty domain to clear it, and skip the https/letsencrypt
requirements when the domain is being removed.
Fixes#4821
The randomize/isolated-deployment volume transform split mount strings
on ':' and kept only the first two segments, so an access mode like
:ro, :z or :Z was silently dropped and read-only mounts became
read-write. Keep the full path+mode remainder when rebuilding the
mount string.
Fixes#4818
API key names longer than 32 characters were rejected by better-auth
with a 400 that surfaced as an opaque INTERNAL_SERVER_ERROR (the generic
"Failed to generate API key" toast). Add a shared name schema (min 1,
max 32, matching better-auth's default maximumNameLength) used by both
the tRPC input and the client form, and surface the limit on the Name
field so users see it before submitting.
Fixes#4798
