DEV Community

CVE Reports
CVE Reports

Posted on Originally published at cvereports.com

CVE-2026-71486: CVE-2026-71486: Uncontrolled Resource Consumption in vLLM Derender Endpoints

#security #cve #cybersecurity

CVE-2026-71486: Uncontrolled Resource Consumption in vLLM Derender Endpoints

Vulnerability ID: CVE-2026-71486
CVSS Score: 4.3
Published: 2026-09-04

CVE-2026-71486 (GHSA-8737-qx52-hjff) is an uncontrolled resource consumption vulnerability in vLLM's derender endpoints before version 0.26.0. An authenticated attacker can supply crafted, deeply nested token structures to exhaust CPU and memory resources, resulting in server denial of service (DoS) or Out of Memory (OOM) crashes. This vulnerability stems from missing input-bounds validation before passing user-supplied structures to computationally intensive decoding routines.

TL;DR

Authenticated attackers can cause a Denial of Service (DoS) or Out of Memory (OOM) crash in vLLM servers by submitting oversized, deeply nested arrays of token IDs or choices to the /v1/completions/derender or /v1/chat/completions/derender endpoints, which the server detokenized without enforcing structure bounds.

⚠️ Exploit Status: POC

Technical Details

  • CWE ID: CWE-400
  • Attack Vector: Network (AV:N)
  • CVSS Score: 4.3 (Medium)
  • EPSS Score: 0.00341 (27.00% Percentile)
  • Impact: Denial of Service (DoS) / OOM Crash
  • Exploit Status: Proof-of-Concept
  • KEV Status: Not Listed

Affected Systems

  • vLLM Engine
  • vLLM Scale-Out API Server
  • vLLM: < 0.26.0 (Fixed in: 0.26.0)

Code Analysis

Commit: 8e61b64

Implement scale-out derender validation limits and token ID guards.

Exploit Details

  • GitHub: Official pull request containing defensive unit tests and regression test vectors.

Mitigation Strategies

  • Upgrade vLLM to version 0.26.0 or newer to enable native input validation bounds.
  • Disable or block access to the /v1/completions/derender and /v1/chat/completions/derender endpoints at the reverse proxy if scale-out derendering is not required.
  • Enforce strict request body size limits (e.g., < 50 KB) on the reverse proxy or API gateway to prevent transmission of massive JSON payloads.
  • Implement rate limiting per API key to mitigate repetitive resource consumption attempts.

Remediation Steps:

  1. Check current vLLM version using the command: pip show vllm
  2. Update vLLM to the patched version: pip install --upgrade vllm>=0.26.0
  3. Verify the deployment version matches the target version using: python -c "import vllm; print(vllm.__version__)"
  4. Configure NGINX or your API gateway to drop requests exceeding a specified size or block the /derender path entirely if unused.

References

Read the full report for CVE-2026-71486 on our website for more details including interactive diagrams and full exploit analysis.

Top comments (0)