CVE-2025-62426

[CVE catalog]

Severity

Medium · CVSS 6.5CVSS:3.1/AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H

Published

Nov 21, 2025, 1:21:29 AM UTC

Updated

Oct 17, 2026, 10:52:20 AM UTC

Source

[NVD] · [CVE record]

EPSS

0.54% · Oct 27, 2026

A vulnerability in vLLM allows an authenticated user to trigger unintended tokenization during chat template processing by supplying crafted chat_template_kwargs to the /v1/chat/completions or /tokenize endpoints. By forcing the server to tokenize very large inputs, an attacker can block the API server’s event loop for extended periods, causing a denial of service and delaying all other requests. vllm: vLLM vulnerable to DoS via large Chat Completion or Tokenization requests with specially crafted `chat_template_kwargs` The flaw is limited to a denial-of-service vector that requires an authenticated user and relies on abusing an optional, non-security-critical parameter (chat_template_kwargs) to force unexpected tokenization during template application, which is computationally expensive but not indicative of data corruption, privilege escalation, or code execution. The attacker cannot break isolation boundaries or execute arbitrary logic—they can only cause the server’s event loop to stall through large crafted inputs, and only if they already have access to the vLLM API. Moreover, the DoS condition is resource-intensive, depends heavily on model size and server configuration, and does not persist once the malicious request completes. Because the impact is bounded to temporary availability degradation without confidentiality or integrity loss, and because exploitation requires legitimate API access and large payloads, this issue aligns with a Moderate severity rather than an Important/High flaw. The CVSS score(s) listed for this vulnerability do not reflect the associated product's status, and are included for informational purposes to better understand the severity of this vulnerability.