| Titel | litellm <= 1.82.5 Incorrect Authorization (CWE-863) |
|---|
| Beschreibung | # Technical Details
A Security Control Bypass vulnerability exists in the `async_pre_call_hook` methods within `enterprise/enterprise_hooks/banned_keywords.py` and `litellm/proxy/hooks/azure_content_safety.py` of litellm.
The application fails to engage content safety guardrails during API invocations generated from standard HTTP completions because the callbacks expect HTTP string states differing from internally formatted async routing definitions.
# Vulnerable Code
File: `enterprise/enterprise_hooks/banned_keywords.py`, `litellm/proxy/hooks/azure_content_safety.py`
Method: `async_pre_call_hook`
Why: Normal HTTP completions (`/v1/chat/completions`) trigger the guardrails pipeline providing `call_type = "acompletion"`. The affected hooks attempt to assert `if call_type == "completion"`, meaning the logic condition evaluates consistently to `False`. The hooks immediately exit without triggering any scanning layers.
# Reproduction
1. Compile `litellm` integrating the `banned_keywords` security hook parameter array and supply a bad keyword list (`["bannedword"]`).
2. Send a POST request to the completions interface submitting `"bannedword"` in the prompt parameters (`curl -X POST http://localhost:4000/v1/chat/completions -d '{"messages":[{"content":"bannedword test"}]}'`).
3. Note the request proxying completely unobstructed directly back from the backend LLM due to the bypassed hook filtering.
# Impact
- Enterprise Banned Keyword rules and Content Filtering routines via Azure are rendered totally inert and easily evaded, destroying the baseline security posture of proxied model communications. |
|---|
| Quelle | ⚠️ https://gist.github.com/YLChen-007/078179224f07cc4e39e4f141a18c817a |
|---|
| Benutzer | Eric-c (UID 96848) |
|---|
| Einreichung | 23.04.2026 10:08 (vor 2 Monaten) |
|---|
| Moderieren | 20.06.2026 19:12 (2 months later) |
|---|
| Status | Akzeptiert |
|---|
| VulDB Eintrag | 372559 [BerriAI litellm bis 1.82.5 Completions Interface banned_keywords.py async_pre_call_hook prompt erweiterte Rechte] |
|---|
| Punkte | 20 |
|---|