← All CVEs

CVE-2026-41523

HIGH 7.5

Published 2026-06-22 · Last modified 2026-06-30

vLLM is an inference and serving engine for large language models (LLMs). Prior to 0.22.0, an assert-based security check in vLLM's activation function loading allows any unauthenticated attacker to achieve arbitrary code execution on the server by publishing a malicious HuggingFace model, when vLLM runs in Python optimized mode (python -O or PYTHONOPTIMIZE=1). This vulnerability is fixed in 0.22.0.

ELEVATED IMPACT

Severe if exploited (CVSS 7.5), but no known exploitation and low modeled probability. Patch on a normal cadence.

Exploitation likelihood

0.5%chance of exploitation in 30 days · 37th percentile

○ In CISA KEV ○ Public exploit / PoC

Impact if exploited

7.5CVSS 3.1 · HIGH

  • ConfidentialityHigh
  • IntegrityHigh
  • AvailabilityHigh

What an attacker needs

  • Access: Reachable over the network — no local access needed
  • Privileges: No account or privileges required
  • User interaction: A user must take an action (click / open a file)
  • Complexity: Needs a race window or specific setup

✓ lowers the bar for an attacker · ⚠ raises it

Affected

Vendors Vllm Project Red Hat

Products Vllm Red Hat Ai Inference Server Red Hat Enterprise Linux Ai (Rhel Ai) 3 Red Hat Openshift Ai (Rhoai)

Weakness (CWE)

  • CWE-94: Code injection
  • CWE-617

CVSS vector

CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:U/C:H/I:H/A:H

Sources: NVD · CVE.org · EPSS