All vulnerabilities

Servers and applications

CVE-2026-93592

Denial of service: a negative token id in `/v1/embeddings` or `/pooling` input kills the vLLM engine via a CUDA device-side assertion

Severity
High 8.7
Published
3 Sept 2026
Last changed
30 Sept 2026
Added to UpHeron
3 Nov 2026
Fixed in
0.28.0
Exploit chance
0.5% in the next 30 daysHigher than 46% of CVEs. EPSS by FIRST, 19 Nov 2026
Weakness
CWE-129
Attack
Over the network · No login needed · No user action needed
Attacker can
Take it down
CVSS vector
CVSS:4.0/AV:N/AC:L/AT:N/PR:N/UI:N/VC:N/VI:N/VA:H/SC:N/SI:N/SA:N/E:X/CR:X/IR:X/AR:X/MAV:X/MAC:X/MAT:X/MPR:X/MUI:X/MVC:X/MVI:X/MVA:X/MSC:X/MSI:X/MSA:X/S:X/AU:X/R:X/V:X/RE:X/U:X
Also known as
GHSA-25q3-v2hm-8vpf
Sources
NVD, Project advisories (GitHub)

Description

A single request carrying a negative token id in the token-id form of the input field of /v1/embeddings or /pooling terminates the vLLM engine. The request is answered HTTP 500 and the server then serves nothing further: /health refuses the connection and every subsequent request from every client fails until the process is restarted.

No authentication, no prior request, and no unusual value are required. -1 is enough.

/v1/embeddings and /pooling accept token ids directly from the caller (input: list[int] | list[list[int]] | str | list[str] on CompletionRequestMixin, in vllm/entrypoints/pooling/base/protocol.py). The engine validated only the upper bound of those ids against the vocabulary, in InputProcessor._validate_model_input:

There was no corresponding lower-bound check, so a negative id passed validation and was used directly as an index into the embedding table on the GPU. That trips a device-side assertion inside PyTorch's index kernel:

From vllm-project/vllm's security advisory. Shortened; the full text is there.

Workarounds

For deployments that cannot upgrade to 0.28.0:

  • Reject request bodies containing negative integers in input at a reverse proxy or API gateway in front of vLLM.
  • Do not expose the token-id form of the input; accept only string input from untrusted callers.
  • Restrict network access to /v1/embeddings and /pooling.

Process supervision is not a workaround: the request can simply be repeated.

Workarounds as given by vllm-project/vllm's security advisory.

Affected software

SoftwareAffected versions
vLLMServers and applications · cpe:a:vllm:vllm
  • Every version before 0.28.0
vLLMServers and applications · github:vllm-project/vllm
  • Every version before 0.28.0

About vLLM

open-source software for large language model inference

From Wikidata, CC0 1.0.
Docker Hub pulls
44,195,225Docker Hub, 15 Nov 2026
GitHub stars
96,943GitHub, 15 Nov 2026
Forks
24,836GitHub, 15 Nov 2026
Latest release
v0.33.0rc3GitHub, 15 Nov 2026
Released
2 Nov 2026GitHub
License
Apache Software License 2.0Wikidata
First published on GitHub
9 Feb 2023GitHub

References

  • NVD https://nvd.nist.gov/vuln/detail/CVE-2026-93592
  • CVE.org https://www.cve.org/CVERecord?id=CVE-2026-93592
  • GitHub Advisory Database https://github.com/advisories/GHSA-25q3-v2hm-8vpf
  • github.com https://github.com/vllm-project/vllm/security/advisories/GHSA-25q3-v2hm-8vpf