CVE-2026-34756 vllm-project CVE debrief

Who should care

Users of vLLM OpenAI-compatible API server versions from 0.1.0 to before 0.19.0 should be aware of this vulnerability and take steps to mitigate it. This includes updating to version 0.19.0 or later, monitoring API server logs for suspicious activity, and considering rate limiting on API requests. Organizations using affected versions in production should prioritize patching due to the potential for denial of service attacks.

Technical summary

The vulnerability in vLLM OpenAI-compatible API server versions from 0.1.0 to before 0.19.0 is caused by the lack of upper bound validation on the n parameter in the ChatCompletionRequest and CompletionRequest Pydantic models. An unauthenticated attacker can exploit this by sending a single HTTP request with an extremely large n value. This causes the Python asyncio event loop to block completely and leads to immediate Out-Of-Memory crashes by allocating millions of request object copies in the heap before the request even reaches the scheduling queue. The CVSS score for this vulnerability is 6.5, with a severity rating of MEDIUM. The vulnerability is addressed in version 0.19.0.

Defensive priority

CVE-2026-34756 has a CVSS score of 6.5 and is rated as MEDIUM severity. Defenders should prioritize patching affected systems, as the vulnerability allows for denial of service attacks with a relatively simple HTTP request.

Recommended defensive actions

• Update vLLM OpenAI-compatible API server to version 0.19.0 or later
• Monitor API server logs for suspiciously large n values in requests
• Consider implementing rate limiting on API requests to mitigate potential attacks
• Review and adjust API server configurations to prevent exploitation
• Perform regular vulnerability scans to detect affected systems

Evidence notes

The CVE-2026-34756 vulnerability is documented in the official CVE record and NVD detail pages. Additional information is available from vendor references and source items provided. The vulnerability is caused by a lack of upper bound validation on the n parameter in certain Pydantic models, allowing for denial of service attacks. The vulnerability is addressed in version 0.19.0 of the vLLM OpenAI-compatible API server.

Official resources