Skip to content

GHSA-5pg6-m483-7vrg CVE-2026-55066 GO-2026-6308

Vikunja has cross-tenant IDOR in kanban move-task endpoint via unauthorized body task_id

High Go
Severity
7.1 / 10
High · CVSS 3.1
Exploit chance (EPSS)
0.37%
Chance of exploitation in the next 30 days. Higher than 30.77% of scored vulnerabilities.
Known exploited
No
Not in CISA's catalog
Weakness
CWE-639 Authorization Bypass Through User-Controlled Key
Fix
2.4.0
Upgrade code.vikunja.io/api to this version or later

Description

Summary

The kanban endpoint POST /api/v1/projects/{project}/views/{view}/buckets/{bucket}/tasks moves a task into a bucket. The task is identified by task_id in the request body. The endpoint's authorization check (TaskBucket.CanUpdate) only verifies that the caller may update the project/view/bucket named in the URL — it never checks any permission on task_id.

Any authenticated user can therefore supply another user's task ID (task IDs are a global, sequential integer space) against a kanban bucket in their own project. The server loads that victim task with no authorization check, returns its full contents in the response, and — when the target bucket is a "done" bucket — writes to the victim task's row.

This is the same vulnerability class Vikunja has already remediated for task relations (CVE-2026-33676), task attachments (CVE-2026-33678), task comments (CVE-2026-33313) and CalDAV task read (CVE-2026-35598). TaskBucket is the task-child operation that was missed.


Root cause

1. task_id is body-controlled and never permission-checked

pkg/models/kanban_task_bucket.go:32:

type TaskBucket struct {
    BucketID      int64 `... json:"bucket_id" param:"bucket"`
    TaskID        int64 `... json:"task_id"`              // body-bound only — no param tag
    ProjectViewID int64 `... json:"project_view_id" param:"view"`
    ProjectID     int64 `xorm:"-" json:"-" param:"project"`
    ...
}

The web handler UpdateWeb (pkg/web/handler/update.go) populates the struct via ctx.Bind, which binds both URL path params (param: tags) and the JSON body. BucketID, ProjectViewID, ProjectID come from the trusted URL; TaskID comes entirely from the attacker-controlled body.

2. CanUpdate authorizes the URL, not the task

pkg/models/kanban_task_bucket.go:52:

func (b *TaskBucket) CanUpdate(s *xorm.Session, a web.Auth) (bool, error) {
    bucket := Bucket{ID: b.BucketID, ProjectID: b.ProjectID, ProjectViewID: b.ProjectViewID}
    return bucket.canDoBucket(s, a)
}

canDoBucket (pkg/models/kanban_permissions.go:46) resolves the bucket/view and ends in Project{ID: pv.ProjectID}.CanUpdate(s, a) — a permission check on the project from the URL. b.TaskID is never referenced. The attacker owns that project, so the check passes.

3. The task is loaded and mutated with no authorization

updateTaskBucket (pkg/models/kanban_task_bucket.go:119):

task := &Task{ID: b.TaskID}
err = task.ReadOne(s, a)            // loads ANY task by ID — no permission check

Task.ReadOne (pkg/models/tasks.go:1967) calls GetTaskByIDSimple + addMoreInfoToTasks; it performs no authorization (authorization normally lives in the separate Task.CanRead, which this internal call path bypasses).

  • Read: the fully populated victim task is assigned to b.Task (line 227) and returned by the Update handler in the response "task" field.

  • Write: if the target bucket is the view's done bucket (view.DoneBucketID == b.BucketID && !task.Done, line 141), the handler sets task.Done = true and persists it to the victim's task row:

    _, err = s.Where("id = ?", task.ID).
        Cols("done", "due_date", "start_date", "end_date", "done_at").
        Update(task)
    

Proof of Concept

The attacker is any normal authenticated user. They first create their own kanban project/view/bucket (free for every user), then:

POST /api/v1/projects/{ATTACKER_PROJECT}/views/{ATTACKER_VIEW}/buckets/{ATTACKER_BUCKET}/tasks HTTP/1.1
Host: TARGET
Authorization: Bearer {ATTACKER_JWT}
Content-Type: application/json

{"task_id": {VICTIM_TASK_ID}}

The 200 response body contains the victim task in full under "task" — title, description, dates, assignees, labels, attachment list, reactions. Task IDs are a global sequential counter, so iterating task_id enumerates every task on the instance.

If ATTACKER_BUCKET is the done bucket of ATTACKER_VIEW, the same request also flips the victim task to done (done = true, done_at set).


Impact

Any authenticated low-privilege user can:

  • Read any task on the instance by sequential ID, across every other user, project and organization — a full cross-tenant information disclosure of task titles, descriptions, assignees, labels and attachment metadata.
  • Modify any task's done state, marking arbitrary victims' tasks done (or clearing it) and altering done_at.

Vikunja's permission model is built specifically to isolate projects between users; this endpoint defeats that isolation. It is the same impact and class that warranted CVEs for task relations, attachments and comments.


Suggested fix

In TaskBucket.CanUpdate, after the bucket/project check, also verify the caller's permission on the body-supplied task — mirroring the remediation already applied to task relations and attachments:

task := &Task{ID: b.TaskID}
canUpdateTask, err := task.CanUpdate(s, a)
if err != nil || !canUpdateTask {
    return false, err
}

(Use CanRead if moving a readable-but-not-writable task into a bucket is intended; CanUpdate is the safer default since the operation can change the task's done state.)


References

  • CWE-639 Authorization Bypass Through User-Controlled Key
  • CWE-284 Improper Access Control
  • OWASP A01:2021 Broken Access Control
  • CVE-2026-33676, CVE-2026-33678, CVE-2026-33313, CVE-2026-35598 — the same missing-authorization-on-task-child class, already remediated; this report is the un-remediated TaskBucket sibling.

Additional notes

  • The v2 API is affected too. The same endpoint is exposed under /api/v2/..., and both versions route through the shared model TaskBucket.CanUpdate / updateTaskBucket in pkg/models/kanban_task_bucket.go. A model-level fix closes v1 and v2 simultaneously; the regression test should assert both.

  • Two fix altitudes. The minimal fix checks the body-supplied task_id in TaskBucket.CanUpdate (task.CanUpdate/CanRead). A broader fix makes Task.ReadOne itself permission-aware, which also hardens other internal call paths that rely on it — higher blast radius, weigh accordingly.

  • Side effects of the cross-tenant write confirmed across reports: flipping done rewrites done_at/due_date/start_date/end_date, inserts a task_buckets row, propagates done-state to other kanban views with a done bucket in the victim's project, and triggers updateDone rescheduling for repeating tasks.

Affected versions

Go

PackageVulnerableFixed in
code.vikunja.io/api
< 2.4.0
2.4.0
code.vikunja.io/api
all versions
No fix yet

How to fix

Move to the fixed version or later with your package manager, update the lockfile, and run your tests. If the package only arrives through another dependency, upgrade that dependency first. If none of its releases pulls a fixed version yet, force the version with your package manager's override.

code.vikunja.io/api Go → 2.4.0

go
go get code.vikunja.io/api@v2.4.0 && go mod tidy

Severity breakdown

CVSS 3.1 7.1

CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:L/A:N

Attack vector
Network
Attack complexity
Low
Privileges required
Low
User interaction
None
Scope
Unchanged
Confidentiality
High
Integrity
Low
Availability
None

Weakness

CWE-639: Authorization Bypass Through User-Controlled Key

The system's authorization functionality does not prevent one user from gaining access to another user's data or record by modifying the key value identifying the data.

How to prevent it
  • Architecture and Design: For each and every data access, ensure that the user has sufficient privilege to access the record that is being requested.
  • Architecture and Design, Implementation: Make sure that the key that is used in the lookup of a specific user's record is not controllable externally by the user or that any tampering can be detected.
  • Architecture and Design: Use encryption in order to make it more difficult to guess other legitimate values of the key or associate a digital signature with the key so that the server can verify that there has been no tampering.

References

Timeline

  1. Published
  2. Last updated by the source
  3. Latest EPSS score
AdvisorySeverityWhy it's related
GHSA-m687-p538-r5hp
Vikunja: Permissive Cross-domain Security Policy trusts every localhost origin which should not be trusted
Medium Also affects code.vikunja.io/api
GHSA-fmmf-xq98-g327
Vikunja: Write-level project members can delete admin-tier link shares through an unloaded permission check
Medium Also affects code.vikunja.io/api
GHSA-pjr3-86v4-5p7w
Vikunja: Explicit lower-permission share on a sub-project is silently overridden by an inherited parent permission (broken access control / privilege-management regression in v2.6.0)
Medium Also affects code.vikunja.io/api
GHSA-hjx8-qv73-f7cm
Vikunja: Webhooks and link shares survive every revocation path, so a removed collaborator keeps a live feed
High Also affects code.vikunja.io/api
GHSA-fprf-r6rv-xg99
Vikunja: Saved filter creation with an empty filter string recalculates task positions across all tenants
Medium Also affects code.vikunja.io/api
GHSA-4hv6-xc92-j86g
Vikunja: WebSocket authentication ignores server-side session state, so revoked sessions keep receiving live pushes
High Also affects code.vikunja.io/api
GHSA-w2ch-4xgr-22ww
Vikunja: Task relation deletion does not check read access to the other task, allowing cross-project relation removal
Low Also affects code.vikunja.io/api
GHSA-8wvg-r2j4-3737
Vikunja: Assignee email addresses disclosed to read-only project members via the task assignees endpoint
Medium Also affects code.vikunja.io/api
Is this in your code?

MergeHarbor checks every repository's dependencies against this database, ranks what matters by exploitability and reachability, routes it to the owning team and opens the fix pull request.

Check your repos with MergeHarbor