Code Injection in vLLM - CVE-2026-41523

Vulnerability details

The vulnerability allows a remote attacker to execute arbitrary code.

The vulnerability exists due to code injection in the activation function loading logic in vllm/model_executor/layers/pooler/activations.py when loading a crafted HuggingFace model configuration while Python optimized mode disables the assert-based security check. A remote attacker can publish a malicious model with a crafted config.json to execute arbitrary code.

User interaction is required to load the malicious model, and exploitation requires Python optimized mode and a cross-encoder architecture.

Affected software

How to mitigate CVE-2026-41523

External References

• https://github.com/vllm-project/vllm/security/advisories/GHSA-q8gq-377p-jq3r
• https://github.com/vllm-project/vllm/commit/b3c7ffcab82c2439726f8cb213800f6f38c023d3

Related Security Bulletins

• Multiple vulnerabilities in vLLM