# CVE-2026-116181: vLLM improper input validation in stop_token_ids handling

Source: https://junglewise.ai/threats/cve-2026-116181-vllm-from-0-20-1-before-0-29-0-contains-an-improper-input  
Publisher: Junglewise Threat Intelligence, https://junglewise.ai/threats  
Published: 2026-10-25. Updated: 2026-10-25  
License: CC BY 4.0 with a link back

CVE-2026-116181 · Severity: medium · CVSS 6.5 · Published 2026-10-25

vLLM is a large language model serving framework that handles API requests to generate text. An attacker can crash the service or corrupt GPU memory by sending requests with invalid token IDs that are not validated against the model's vocabulary. This causes service disruption for all users and potential data corruption on GPU.

Technologies: [vLLM Project vllm](https://junglewise.ai/threats/technologies/vllm-project-vllm). Vendors: [vLLM Project](https://junglewise.ai/threats/vendors/vllm-project).

## Quick answers

**Is CVE-2026-116181 exploited in the wild?** No. Junglewise Threat Intelligence has no report of CVE-2026-116181 being exploited in the wild.

**How severe is CVE-2026-116181?** Medium severity, with a CVSS score of 6.5, published 25 October 2026.

**Which products and versions are affected?** vLLM Project vllm 0.20.1 before 0.29.0.

**Is there a fix for CVE-2026-116181?** Yes: fixed in 0.29.0. Junglewise finds every place you run it, drafts the upgrade and applies it once you approve.

**How is CVE-2026-116181 attacked?** It is attacked over the network, in a software library.

**Am I affected by CVE-2026-116181?** Check whether you run vLLM Project vllm at an affected version. Junglewise does that check for you around the clock: connect your code, cloud and domains, and it matches new vulnerabilities like this one against what you actually run and tells you when one reaches you.

## How to fix

- Fixed in 0.29.0

## Technical details

The vulnerability is an improper array index validation (CWE-129) in SamplingParams that accepts any integer for stop_token_ids without checking if the ID exists in the model vocabulary. These unvalidated IDs are used as column indices into the logits tensor; on the V1 runner this triggers an assertion crash, while on the V2 runner it writes out-of-bounds to the logits buffer. The flaw affects OpenAI chat, completion, generate, and batch API endpoints and requires only low-privilege authenticated network access.

## Affected products

- vLLM Project vLLM 0.20.1 before 0.29.0

## Timeline

- 2026-10-09: advisory
- 2026-10-25: patched

## References

- [https://github.com/vllm-project/vllm/security/advisories/GHSA-w9r6-29q4-pvpr](https://github.com/vllm-project/vllm/security/advisories/GHSA-w9r6-29q4-pvpr)
- [https://www.vulncheck.com/advisories/vllm-0.20.1-before-0.29.0-dos-and-gpu-out-of-bounds-write-via-stop-token-ids](https://www.vulncheck.com/advisories/vllm-0.20.1-before-0.29.0-dos-and-gpu-out-of-bounds-write-via-stop-token-ids)

## Related threats

- [vLLM chat template resource exhaustion DoS](https://junglewise.ai/threats/cve-2026-115918-vllm-is-an-inference-and-serving-engine-for-large-language)
- [vLLM off-by-one error in tokenizer max_token_id validation](https://junglewise.ai/threats/cve-2026-115916-vllm-is-an-inference-and-serving-engine-for-large-language)
- [vLLM image format validation bypass in media processing](https://junglewise.ai/threats/cve-2026-115915-vllm-is-an-inference-and-serving-engine-for-large-language)
- [vLLM Phi-4 multimodal audio encoder GPU memory exhaustion](https://junglewise.ai/threats/cve-2026-115913-vllm-is-an-inference-and-serving-engine-for-large-language)
- [vLLM arbitrary code execution in shared-memory multimodal cache](https://junglewise.ai/threats/cve-2026-115912-vllm-is-an-inference-and-serving-engine-for-large-language)