vLLM before 0.29.0 fails to enforce decoder prompt-length validation on the disaggregated serving endpoint /inference/v1/generate. When the request contains a 'features' (multimodal) payload, vllm/entrypoints/serve/disagg/serving.py builds a multimodal EngineInput directly from the caller-supplied token_ids, and GenerateRequest.token_ids (vllm/entrypoints/serve/disagg/protocol.py) is not checked against model_config.max_model_len. For multimodal processors that report skip_prompt_length_check=True (for example Nemotron Parse, Whisper, and FireRedLID), InputProcessor._validate_prompt_len() returns immediately for both encoder and decoder prompts, so an overlong prompt becomes an EngineCoreRequest and reaches the worker input-batch copy into a fixed max_model_len-wide NumPy row. A client able to reach the endpoint on an affected model configuration can therefore submit an overlong token_ids list to trigger a worker failure and denial of service. Fixed in 0.29.0.
| Métrique | Valeur | Interprétation |
|---|---|---|
|
AV (Vecteur d'Attaque)
Attack Vector
|
Network (N)
|
Exploit possible à distance via réseau (HTTP, API, etc.) |
|
AC (Complexité d'Attaque)
Attack Complexity
|
Low (L)
|
Peu de conditions requises pour exploiter |
|
PR (Privilèges Requis)
Privileges Required
|
Low (L)
|
Authentifié avec compte utilisateur basique |
|
UI (Interaction Utilisateur)
User Interaction
|
None (N)
|
Aucune interaction utilisateur nécessaire |
|
S (Périmètre)
Scope
|
Unchanged (U)
|
Impact reste dans le périmètre du composant compromis |
|
C (Impact Confidentialité)
Confidentiality Impact
|
None (N)
|
Aucune divulgation d'informations |
|
I (Impact Intégrité)
Integrity Impact
|
None (N)
|
Aucune modification de données possible |
|
A (Impact Disponibilité)
Availability Impact
|
High (H)
|
Interruption totale du service possible |
vllm before 0.29.0 Denial of Service via Decoder Prompt Length Bypass