CVE-2026-93592
Description
A flaw was found in vLLM. An unauthenticated attacker can exploit an input validation vulnerability by submitting negative token IDs to the /v1/embeddings and /pooling endpoints. This action triggers a CUDA device-side assertion, which poisons the GPU context and causes the engine to crash. Consequently, all subsequent requests will fail until the process is restarted, leading to a Denial of Service (DoS).
Statement
Red Hat has evaluated this vulnerability as having an Important severity because an unauthenticated remote attacker can completely disrupt the inference engine with a single payload. In Red Hat OpenShift AI and Red Hat AI Inference Server deployments exposing embedding or pooling models, processing negative token inputs causes an unrecoverable GPU context failure. This fault halts the entire serving process and prevents subsequent client requests from being handled until the container or process is restarted.
Mitigation
Restrict access to the inference service or filter requests at an upstream reverse proxy or API gateway:
1. Limit network access: Deploy firewall rules or OpenShift NetworkPolicies so that the inference server ports and endpoints are accessible only by authenticated, trusted internal clients.
2. Ingress request validation: Configure an API gateway or reverse proxy placed in front of the model server to inspect request payloads for the `/v1/embeddings` and `/pooling` endpoints, rejecting requests where the `input` array contains negative numbers, or restricting untrusted clients to submitting text prompt strings rather than raw token lists.
Caveats: Inspecting request bodies at an intermediary gateway layer may introduce slight processing latency. Modifying reverse proxy rules or applying network policies may require a service configuration reload or restart, which can temporarily disrupt in-flight connections.
Common Vulnerability Scoring System (CVSS) Score Details
Info alert:Important note
CVSS scores for open source components depend on vendor-specific factors (e.g. version or build chain). Therefore, Red Hat's score and impact rating can be different from NVD and other vendors. Red Hat remains the authoritative CVE Naming Authority (CNA) source for its products and services (see Red Hat classifications).
CVSS v3 Score Breakdown
| Red Hat | NVD | cve.org | |
|---|---|---|---|
| Base Score | 7.5 | N/A | 7.5 |
| Attack Vector | Network | N/A | Network |
| Attack Complexity | Low | N/A | Low |
| Privileges Required | None | N/A | None |
| User Interaction | None | N/A | None |
| Scope | Unchanged | N/A | Unchanged |
| Confidentiality | None | N/A | None |
| Integrity Impact | None | N/A | None |
| Availability Impact | High | N/A | High |
Vector
Red Hat: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
cve.org: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:H
Understanding the Weakness (CWE)
Availability
Technical Impact: DoS: Crash, Exit, or Restart
An attacker that can trigger an assert statement can still lead to a denial of service if the relevant code can be triggered by an attacker, and if the scope of the assert() extends beyond the attacker's own session.
Frequently Asked Questions
Not sure what something means? Check out our Security Glossary.
Want to get errata notifications? Sign up here.
