CVE-2026-91984 — api

Summary

The task-position endpoint authorizes only the task side of the write: TaskPosition.CanUpdate delegates to Task.CanUpdate (write access to the task's own project) and the request body's project_view_id is never validated to belong to the task's project, nor is any access to that view required. Any authenticated user with a single writable task of their own can persist (task_id, project_view_id, position) rows into any other tenant's project view (view IDs are small sequential integers and enumerable). Low position values (below MinPositionSpacing, 0.01) enter the recalculation branch, but RecalculateTaskPositions aborts on its own project read-access check before recalculating and the whole transaction rolls back, so no cross-tenant recalculation actually executes — only the plain row insert (position ≥ 0.01) persists. This is the same root-cause pattern as GHSA-569v-q83c-3j3g (kanban bucket relocation via project_view_id mass-assignment, fixed with a dual-side check for buckets in 2.4.0) surviving in the sibling position endpoint. Verified on Vikunja 2.5.0.

Details

Affected endpoints (both verified):

• v1: POST /api/v1/tasks/{id}/position
• v2: PUT /api/v2/tasks/{id}/position

Root cause in pkg/models/task_position.go:

• CanUpdate (lines 61-64) checks only Task.CanUpdate — i.e. the caller's write access to the task's own project.
• updateTaskPosition (lines 174-232) upserts (task_id, project_view_id, position) directly; there is no check that the view belongs to the task's project and no access check on the view's project.
• ProjectViewID is body-bindable (json:"project_view_id", no readOnly/param restriction, line 42).

Contrast with the fixed sibling: the bucket-move endpoint (POST /projects/{project}/views/{view}/buckets/{bucket}/tasks, pkg/models/kanban_task_bucket.go:53-68) validates both the bucket and the task after the GHSA-5pg6/GHSA-569v family fixes — the position endpoint was never given the equivalent view-side check.

Observed behavior (two independent verification runs): an attacker with zero access to the victim project receives 200 {"task_id": ..., "project_view_id": <victim view>, "position": ...} on both v1 and v2, and the row is persisted (verified directly in the task_positions table: rows for the victim view went 0 to 1). Positioning the victim's own task is correctly refused with 403, isolating the missing view-side check. The victim's listings do not surface the foreign task (no confidentiality impact demonstrated), and a low-position (< 0.01) injection enters the recalculation branch but RecalculateTaskPositions returns 403 (getRelevantProjectsFromCollection → CanRead on the victim project) before any recalculation runs, rolling the whole request back — so the low-position variant persists nothing and never rewrites the victim's ordering.

PoC

Steps use the owner of the victim project (token $OWNER) and an attacker with only their own project (token $ATTACKER). All requests were executed against a local Vikunja 2.5.0 instance.

1. Owner creates a private project (default views included) and lists its views to obtain a victim view ID:

curl -X PUT "$BASE/api/v1/projects" -H "Authorization: Bearer $OWNER" \
  -H 'Content-Type: application/json' -d '{"title":"victim"}'
# -> {"id":200,...}

curl "$BASE/api/v1/projects/200/views" -H "Authorization: Bearer $OWNER"
# -> [{"id":771,"view_kind":"kanban",...}, ...]   # remember VIEW=771

1. Attacker creates their own project and a task in it:

curl -X PUT "$BASE/api/v1/projects" -H "Authorization: Bearer $ATTACKER" \
  -H 'Content-Type: application/json' -d '{"title":"attacker"}'
# -> {"id":201,...}

curl -X PUT "$BASE/api/v1/projects/201/tasks" -H "Authorization: Bearer $ATTACKER" \
  -H 'Content-Type: application/json' -d '{"title":"evil"}'
# -> {"id":105,...}

1. Baseline — attacker positions their task in their own view (succeeds, expected):

curl -X POST "$BASE/api/v1/tasks/105/position" -H "Authorization: Bearer $ATTACKER" \
  -H 'Content-Type: application/json' -d '{"project_view_id": <OWN_VIEW>, "position": 100}'
# -> 200

1. Mutation — attacker writes their task's position into the victim's view (no access to it whatsoever):

curl -X POST "$BASE/api/v1/tasks/105/position" -H "Authorization: Bearer $ATTACKER" \
  -H 'Content-Type: application/json' -d '{"project_view_id": 771, "position": 100}'
# -> 200 {"task_id":105,"project_view_id":771,"position":100}

curl -X PUT "$BASE/api/v2/tasks/105/position" -H "Authorization: Bearer $ATTACKER" \
  -H 'Content-Type: application/json' -d '{"project_view_id": 771, "position": 101}'
# -> 200 (v2 twin behaves identically)

1. Control — attacker positioning the victim's own task is correctly refused (the task-level gate works; only the view side is missing):

curl -X POST "$BASE/api/v1/tasks/<VICTIM_TASK>/position" -H "Authorization: Bearer $ATTACKER" \
  -H 'Content-Type: application/json' -d '{"project_view_id": 771, "position": 1}'
# -> 403 Forbidden

1. Persistence proof — the cross-tenant row exists in the database:

docker exec <db> psql -U vikunja -d vikunja -t -A -c \
  "SELECT task_id, position FROM task_positions WHERE project_view_id=771 AND task_id=105;"
# -> 105|100        (row for the attacker's task inside the victim's view namespace)

Observed result: steps 4 and 6 succeed — a task_positions row (attacker_task, victim_view) is created and persisted; step 5 is refused with 403. Note: "position": 1 is above MinPositionSpacing (0.01), so it never enters the recalculation branch — it is a plain upsert. A position < 0.01 does enter the branch but is refused with 403 and rolled back (see Impact).

Impact

Impact summary: a genuine broken-object-level-authorization defect (CWE-639) with no demonstrated confidentiality, integrity, or availability impact. The injected rows are invisible to the victim (view listings filter by project), order-preserving (any position collision is respaced between the same neighbours), and self-healing (any legitimate recalculation of the view deletes and rebuilds all its position rows). No read access, no privilege gain, no data destruction, and no usable DoS. Rated low — the fix is defense-in-depth, and its main value is preventing this wrong-object authorization from becoming a real cross-tenant leak should any future read path join task_positions without re-checking task-project access.

Broken object-level authorization on the write side (CWE-639, CWE-863: the authorization decision uses one resource dimension — the task — while the write targets another — the view). Any authenticated user holding a single writable task can persist rows into any other tenant's view state instance-wide (view IDs sequential and enumerable). The recalculation/lock path is not usable cross-tenant: a low position (< 0.01) enters the branch but RecalculateTaskPositions fails its own project read-access check and rolls back before recalculating (on MySQL/PostgreSQL a FOR UPDATE lock on the view row is taken one statement earlier in the same transaction, but released immediately on that rollback; SQLite takes no lock), so there is no meaningful availability angle. No confidentiality breach was demonstrated (victim listings filter by project), and the pollution is recoverable. Suggested fix: in CanUpdate or before the upsert, load the view by tp.ProjectViewID and require view.ProjectID == task.ProjectID (and/or caller access to the view's project), mirroring the bucket-side fix for GHSA-569v.