UTC
Medium

Disaggregated generate skips decoder prompt-length validation for some multimodal processors

GHSA-3mqx-f33v-vgp9 · published 2026-09-12

Severity
Medium
CVSS score
6.5
Chance of exploitation (EPSS)
No CVE
Fixed in
0.29.0
What it is

An API client that can call /inference/v1/generate against an affected skip-check multimodal deployment can submit a malformed request that fails inside worker input batching before model execution. This is an availability issue: the request should be rejected as invalid input, but the skip-check configuration allows it to reach worker-side batch construction. Suggested CWE: CWE-400: Uncontrolled Resource Consumption CWE-754: Improper Check for Unusual or Exceptional Conditions Suggested CVSS v3.1 when API-key authentication is configured: Score: 6.5 Medium. If the server is exposed without API-key authentication, the deployment-specific vector would use PR:N.

What to do

Upgrade vllm to 0.29.0 or later. If you start it from an MCP config with npx or uvx, pin that version so an old cached copy is not used.

Do not let the multimodal encoder-padding skip flag disable decoder prompt-length validation. A minimal fix is to split the check by prompt type: keep the encoder-specific skip behavior only where the padded encoder prompt requires it; always enforce decoder prompt_len > max_model_len and generate prompt_len == max_model_len rejection before constructing EngineCoreRequest; add regression coverage for /inference/v1/generate with features present, explicit sampling_params.max_tokens, a skip-check multimodal model configuration, and len(token_ids) == max_model_len + 1; assert a controlled 400-style prompt-length error and that the worker input batch is not reached.

Affected
PackageRegistryAffectedFixed in
vllmPyPI< 0.29.00.29.0

The impact, patch and workaround text comes from the advisory, shortened (published by vLLM). Proof-of-concept and reproduction sections and code are left out; the original advisory has them.

Sources: GitHub Advisory Database, reviewed and malware advisories (CC BY 4.0, https://github.com/advisories); advisories published by Cursor (https://github.com/cursor/cursor/security/advisories); advisories published by OpenClaw (https://github.com/openclaw/openclaw/security/advisories); advisories published by vLLM (https://github.com/vllm-project/vllm/security/advisories); advisories published by Open WebUI (https://github.com/open-webui/open-webui/security/advisories); advisories published by n8n (https://github.com/n8n-io/n8n/security/advisories); advisories published by Flowise (https://github.com/FlowiseAI/Flowise/security/advisories); advisories published by Langflow (https://github.com/langflow-ai/langflow/security/advisories); advisories published by the MCP Python SDK project (https://github.com/modelcontextprotocol/python-sdk/security/advisories); advisories published by MLflow (https://github.com/mlflow/mlflow/security/advisories); CISA Known Exploited Vulnerabilities catalog (CC0). Data as of 2026-11-22. ShieldMap describes how each problem works and how to fix it; it does not publish exploit code. Report a correction to [email protected].