CVE-2026-115916: vLLM tokenizer off-by-one engine termination

# CVE-2026-115916: vLLM tokenizer off-by-one engine termination

https://gcdradar.vercel.app/incidents/cve-2026-115916-vllm-tokenizer-off-by-one-engine-termination · AIR-2026-1543 · closed 2026-10-26

- Lane: Patch now
- Severity: 2 (Disclosed vulnerability): Disclosed vulnerability with a fix and no exploitation reported; CVSS 6.5.
- Fix status: Patched
- First reported: 2026-10-26
- AI role: Attack on AI
- Products: vLLM
- CVE ids: CVE-2026-115916
- CVSS: 6.5

## Summary

An off-by-one bound validation error in vLLM tokenizer handling allows pre-tokenized prompt IDs to trigger device-side assertions and terminate the engine. The issue is fixed in version 0.32.0.

## What to do

Update to vLLM version 0.32.0.

## Affected and fixed versions

- Affected: vLLM < 0.32.0
- Fixed in: vLLM 0.32.0

## Latest updates

- 2026-10-28: GitHub advisory published (https://github.com/advisories/GHSA-q97c-p368-75c4)
- 2026-10-26: CVE-2026-115916: vLLM is an inference and serving engine for large language models. (https://nvd.nist.gov/vuln/detail/CVE-2026-115916)

## Sources

- 2026-10-26 NVD: CVE-2026-115916: vLLM is an inference and serving engine for large language models. https://nvd.nist.gov/vuln/detail/CVE-2026-115916
- 2026-10-28 GitHub Advisory Database: CVE-2026-115916: vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS) https://github.com/advisories/GHSA-q97c-p368-75c4

Verify against the vendor's advisory before changing production. Data CC BY-SA 4.0, https://gcdradar.vercel.app.