GHSA-RCVP-6FGW-C7FH
Vulnerability from github – Published: 2026-05-08 19:52 – Updated: 2026-05-15 23:53Ollama Model Access Control Bypass via /api/generate, /api/embed, /api/embeddings, and /api/show
Affected Component
Ollama proxy endpoints missing model access control:
- backend/open_webui/routers/ollama.py (lines 955-995, generate_completion)
- backend/open_webui/routers/ollama.py (lines 835-881, embed)
- backend/open_webui/routers/ollama.py (lines 891-937, embeddings)
- backend/open_webui/routers/ollama.py (lines 791-820, show_model_info)
Affected Versions
Current main branch (commit 6fdd19bf1) and likely all versions with Ollama model access control support.
Description
Four Ollama proxy endpoints accept any model name from the user and forward the request to the Ollama backend without checking whether the user is authorized to access that model. These endpoints only require get_verified_user (any authenticated non-pending user) and validate that the model exists in the full unfiltered model list, but never check AccessGrants.has_access().
This is in direct contrast with the /ollama/api/chat endpoint (line 1101-1122) which correctly validates model access grants and returns 403 for unauthorized users:
# /api/chat (line 1101-1122) — CORRECTLY checks access
if not bypass_filter and user.role == 'user':
user_group_ids = {group.id for group in Groups.get_groups_by_member_id(user.id)}
if not (
user.id == model_info.user_id
or AccessGrants.has_access(
user_id=user.id, resource_type='model',
resource_id=model_info.id, permission='read',
user_group_ids=user_group_ids,
)
):
raise HTTPException(status_code=403, detail='Model not found')
# /api/generate (line 955-995) — NO access check at all
# /api/embed (line 835-881) — NO access check at all
# /api/embeddings (line 891-937) — NO access check at all
# /api/show (line 791-820) — NO access check at all
CVSS 3.1 Breakdown
| Metric | Value | Rationale |
|---|---|---|
| Attack Vector | Network (N) | Exploited remotely via API calls |
| Attack Complexity | Low (L) | Single API call with a known model name |
| Privileges Required | Low (L) | Requires any authenticated user account |
| User Interaction | None (N) | No victim interaction required |
| Scope | Unchanged (U) | Impact within the Ollama model access boundary |
| Confidentiality | Low (L) | /api/show exposes restricted model details including system prompts and parameters |
| Integrity | None (N) | No data modification |
| Availability | Low (L) | Unauthorized consumption of GPU/compute resources on restricted models |
Attack Scenario
- Admin configures model access control, restricting
llama3:70bto the "ML Engineers" group. Regular user Alice is only authorized forllama3:8b. - Alice knows the restricted model name (model names are predictable —
llama3:70b,mistral:latest, etc.). - Alice calls the unprotected endpoints directly: ```bash # Run completions on restricted model curl -X POST /ollama/api/generate \ -H "Authorization: Bearer " \ -d '{"model": "llama3:70b", "prompt": "..."}'
# View restricted model details and system prompt curl -X POST /ollama/api/show \ -H "Authorization: Bearer " \ -d '{"model": "llama3:70b"}'
# Generate embeddings with restricted model curl -X POST /ollama/api/embed \ -H "Authorization: Bearer " \ -d '{"model": "llama3:70b", "input": "..."}' ``` 4. All requests succeed and are proxied to Ollama without any access control check.
Impact
- Model access control is silently ineffective for four out of five Ollama proxy endpoints
- Unauthorized users can consume GPU/compute resources on restricted models (cost and capacity impact in multi-user deployments)
/api/showexposes restricted model configurations including system prompts, parameters, templates, and license information- Admins have a false sense of security — access restrictions appear to work via the main chat interface but are trivially bypassed via direct API calls
Preconditions
- Ollama must be configured as a backend
- Admin must have configured model access control (not using
BYPASS_MODEL_ACCESS_CONTROL=true) - Attacker must know the restricted model name (model names follow predictable conventions)
{
"affected": [
{
"database_specific": {
"last_known_affected_version_range": "\u003c= 0.8.12"
},
"package": {
"ecosystem": "PyPI",
"name": "open-webui"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "0.9.0"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-44563"
],
"database_specific": {
"cwe_ids": [
"CWE-862"
],
"github_reviewed": true,
"github_reviewed_at": "2026-05-08T19:52:42Z",
"nvd_published_at": "2026-05-15T20:16:48Z",
"severity": "MODERATE"
},
"details": "# Ollama Model Access Control Bypass via /api/generate, /api/embed, /api/embeddings, and /api/show\n\n## Affected Component\n\nOllama proxy endpoints missing model access control:\n- `backend/open_webui/routers/ollama.py` (lines 955-995, `generate_completion`)\n- `backend/open_webui/routers/ollama.py` (lines 835-881, `embed`)\n- `backend/open_webui/routers/ollama.py` (lines 891-937, `embeddings`)\n- `backend/open_webui/routers/ollama.py` (lines 791-820, `show_model_info`)\n\n## Affected Versions\n\nCurrent main branch (commit `6fdd19bf1`) and likely all versions with Ollama model access control support.\n\n## Description\n\nFour Ollama proxy endpoints accept any model name from the user and forward the request to the Ollama backend without checking whether the user is authorized to access that model. These endpoints only require `get_verified_user` (any authenticated non-pending user) and validate that the model exists in the full unfiltered model list, but never check `AccessGrants.has_access()`.\n\nThis is in direct contrast with the `/ollama/api/chat` endpoint (line 1101-1122) which correctly validates model access grants and returns 403 for unauthorized users:\n\n```python\n# /api/chat (line 1101-1122) \u2014 CORRECTLY checks access\nif not bypass_filter and user.role == \u0027user\u0027:\n user_group_ids = {group.id for group in Groups.get_groups_by_member_id(user.id)}\n if not (\n user.id == model_info.user_id\n or AccessGrants.has_access(\n user_id=user.id, resource_type=\u0027model\u0027,\n resource_id=model_info.id, permission=\u0027read\u0027,\n user_group_ids=user_group_ids,\n )\n ):\n raise HTTPException(status_code=403, detail=\u0027Model not found\u0027)\n\n# /api/generate (line 955-995) \u2014 NO access check at all\n# /api/embed (line 835-881) \u2014 NO access check at all\n# /api/embeddings (line 891-937) \u2014 NO access check at all\n# /api/show (line 791-820) \u2014 NO access check at all\n```\n\n## CVSS 3.1 Breakdown\n\n| Metric | Value | Rationale |\n|--------|-------|-----------|\n| Attack Vector | Network (N) | Exploited remotely via API calls |\n| Attack Complexity | Low (L) | Single API call with a known model name |\n| Privileges Required | Low (L) | Requires any authenticated user account |\n| User Interaction | None (N) | No victim interaction required |\n| Scope | Unchanged (U) | Impact within the Ollama model access boundary |\n| Confidentiality | Low (L) | `/api/show` exposes restricted model details including system prompts and parameters |\n| Integrity | None (N) | No data modification |\n| Availability | Low (L) | Unauthorized consumption of GPU/compute resources on restricted models |\n\n## Attack Scenario\n\n1. Admin configures model access control, restricting `llama3:70b` to the \"ML Engineers\" group. Regular user Alice is only authorized for `llama3:8b`.\n2. Alice knows the restricted model name (model names are predictable \u2014 `llama3:70b`, `mistral:latest`, etc.).\n3. Alice calls the unprotected endpoints directly:\n ```bash\n # Run completions on restricted model\n curl -X POST /ollama/api/generate \\\n -H \"Authorization: Bearer \u003calice_token\u003e\" \\\n -d \u0027{\"model\": \"llama3:70b\", \"prompt\": \"...\"}\u0027\n\n # View restricted model details and system prompt\n curl -X POST /ollama/api/show \\\n -H \"Authorization: Bearer \u003calice_token\u003e\" \\\n -d \u0027{\"model\": \"llama3:70b\"}\u0027\n\n # Generate embeddings with restricted model\n curl -X POST /ollama/api/embed \\\n -H \"Authorization: Bearer \u003calice_token\u003e\" \\\n -d \u0027{\"model\": \"llama3:70b\", \"input\": \"...\"}\u0027\n ```\n4. All requests succeed and are proxied to Ollama without any access control check.\n\n## Impact\n\n- Model access control is silently ineffective for four out of five Ollama proxy endpoints\n- Unauthorized users can consume GPU/compute resources on restricted models (cost and capacity impact in multi-user deployments)\n- `/api/show` exposes restricted model configurations including system prompts, parameters, templates, and license information\n- Admins have a false sense of security \u2014 access restrictions appear to work via the main chat interface but are trivially bypassed via direct API calls\n\n## Preconditions\n\n- Ollama must be configured as a backend\n- Admin must have configured model access control (not using `BYPASS_MODEL_ACCESS_CONTROL=true`)\n- Attacker must know the restricted model name (model names follow predictable conventions)",
"id": "GHSA-rcvp-6fgw-c7fh",
"modified": "2026-05-15T23:53:14Z",
"published": "2026-05-08T19:52:42Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/open-webui/open-webui/security/advisories/GHSA-rcvp-6fgw-c7fh"
},
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-44563"
},
{
"type": "PACKAGE",
"url": "https://github.com/open-webui/open-webui"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:L/I:N/A:L",
"type": "CVSS_V3"
}
],
"summary": "Open WebUI\u0027s Ollama Model Access Control Bypass via /api/generate, /api/embed, /api/embeddings, and /api/show"
}
Drag nodes to arrange them. Double-click a related vulnerability to open it, or a source to visit its link. The Physics button on the left tunes the layout and the simulation; the legend at the bottom left hides or shows nodes by kind and by sighting type.
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
Vulnerability timeline
Sightings and recorded dates, grouped by day · UTC
How durations are calculated
Following the VTEM framework (Black Hat USA 2015), time to patch starts when the vendor or a coordinator was informed, total time to patch at discovery, and disclosure to patch at public disclosure; vendor response time runs from the report to its acknowledgement. Time to exploit and time unpatched start at the earliest discovery or report, falling back to public disclosure. Unpatched time ends at the first recorded patch, or the calculation time when no patch is recorded. Exploitation to patch shows the interval between those two events; a published proof of concept is a public exploit, not exploitation.
Each result shows its source dates. A negative interval means the end event came first; when date-only inputs leave the order open, the result says so. Date-only inputs are approximate. The earliest evidence wins: KEV report dates and sightings are labelled as such. Records describing another vulnerability are listed but never calculated from, and a PoC alone does not establish exploitation. Missing patch evidence does not prove that no patch exists. Filters below change the event list only.
Loading timeline…
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.
Related by attack behaviour
Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.