The vLLM project, an open-source library widely used to run large language model inference at scale, released version 0.32.0 on November 17, 2026 patching three security advisories rated from CVSS 7.8 to 9.1. All three vulnerabilities are reachable without authentication on default deployments. Teams running vLLM as a production inference server should upgrade immediately.
GHSA-542m: unsafe deserialization (CVSS 9.1)
The highest-severity issue is an unsafe deserialization vulnerability in vLLM's distributed inference coordinator. The coordinator uses Python's pickle protocol to exchange tensor metadata between worker processes. An attacker with network access to the coordinator port can submit a malicious pickled object that executes arbitrary code in the vLLM worker process. The coordinator port is not authenticated by default and is typically exposed on the internal cluster network.
GHSA-6757: SSRF in model download endpoint (CVSS 8.3)
The second issue is a server-side request forgery vulnerability in the model download endpoint. When a client requests that vLLM download a model from a URL, the server fetches the URL without restricting the destination to trusted domains. An attacker can supply an internal network URL to probe or interact with internal services that the vLLM host can reach, including cloud metadata endpoints at 169.254.169.254, internal registries, and other backend services.
GHSA-h3rc: path traversal in LoRA adapter loader (CVSS 7.8)
The third issue is a path traversal in the LoRA adapter loader. LoRA adapters are small fine-tuning weight files loaded on top of a base model. The loader does not sanitize the adapter file path, allowing an attacker to load arbitrary files from the host filesystem as a model adapter. In some configurations this causes the files to be parsed as tensor data in ways that lead to code execution.
Upgrade path
Upgrade to vLLM 0.32.0 from PyPI. The release notes confirm backward compatibility with existing model deployments. If immediate upgrade is not possible, restrict network access to the vLLM coordinator port and the model download API to trusted internal clients only as a temporary mitigation. This addresses GHSA-542m and partially GHSA-6757, but does not address GHSA-h3rc.
Durable lesson
AI serving infrastructure inherits all the attack surface of a traditional web service: unauthenticated API endpoints, network-accessible internal services, and file system access. The assumption that a vLLM deployment is internal and therefore safe is the same assumption that made internal Jenkins servers and Kubernetes dashboards easy targets. Apply the same access controls you would apply to any externally exposed service: authentication on all endpoints, network segmentation, and input validation at the application boundary.