Race condition in vLLM - #VU139882

Vulnerability details

The vulnerability allows a remote attacker to cause a denial of service or memory corruption.

The vulnerability exists due to a race condition in safe_load_prompt_embeds in vllm/renderers/embed_utils.py when processing concurrent prompt_embeds parts in a /v1/chat/completions request. A remote attacker can send specially crafted concurrent prompt embedding parts to cause a denial of service or memory corruption.

Exploitation requires the enable_prompt_embeds feature to be enabled. Authentication is optional in the stock server and depends on whether an API key is configured.

Affected software

Remediation

External References

• https://github.com/vllm-project/vllm/security/advisories/GHSA-pr7f-p5mw-fc87
• https://github.com/vllm-project/vllm/commit/14043dfecd35dd2f12b4d51eb9fa166184a0ca0f

Related Security Bulletins

• Multiple vulnerabilities in vLLM