UTC
Medium

Remote media is fetched and fully materialized before the documented media size / item limits are enforced — remote pre-inference memory and

GHSA-p6g9-7v3x-m8mv · published 2026-09-12

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

A single client that can reach an affected media ingress path can force the shared serving process (or batch runner) to allocate memory and consume outbound bandwidth proportional to an attacker-chosen compressed body size (Sites 1, 2, 3) or media item count (Site 4) before the request is rejected. Repeated or concurrent requests compound the pressure, up to out-of-memory of a single process. The documented VLLM_MAX_AUDIO_CLIP_FILESIZE_MB / --limit-mm-per-prompt controls give operators a false sense of protection on these surfaces because they run too late or not at all. …

What to do

No fixed release is listed yet for vllm. Limit what it can reach, or stop using it until a fix ships, and follow the source advisory for updates.

Enforce a shared compressed-byte / item-count budget at the acquisition boundary before vLLM creates a full attacker-sized bytes object or spawns per-part fetch work, then keep the existing decode-stage limits as second-stage guards. Add a max_bytes parameter to HTTPConnection.get_bytes / async_get_bytes: honor Content-Length when present, stream in bounded (e.g. 64 KiB) chunks, and abort once the cumulative total exceeds the budget. Only cache or hand off to a decoder after the budget check passes. For inline data: / input_audio payloads, reject from the encoded length ( ) before b64decode, with a post-decode length backstop. …

Affected
PackageRegistryAffectedFixed in
vllmPyPI<= 0.29.0No fix yet

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].