CVE-2026-105755

vLLM: Flash late-interaction scoring caches query embeddings under a caller-controlled request id — cross-request integrity break and induced errors on `/score` and `/rerank`

Published Oct 5, 2026 CVSS 3.1 · 4.2 Medium EPSS < 1% KEV · Not listed
Description
vLLM is an inference and serving engine for large language models. Prior to 0.30.0, flash late-interaction scoring at the /score and /rerank endpoints derives each worker's query_key value from the caller-controlled X-Request-Id header. A concurrent request that reuses a victim's identifier can overwrite the cached query embedding so the victim's documents are scored against the attacker's query, and shared use counters can also cause a late-interaction cache-miss error. This issue is fixed in version 0.30.0.

Analysis

AI-generated

Technical impact

vLLM is an inference engine for large language models. The flaw allows a caller‑supplied X-Request-Id header to be used as a cache key for query embeddings at the /score and /rerank endpoints. A malicious user can send a request that shares a victim’s request identifier and overwrite the cached embedding, causing the victim’s documents to be evaluated against the attacker’s query. This defect also can trigger late‑interaction cache‑miss errors that may disrupt scoring operation. The impact is a loss of integrity for query scoring and potential service disruption, but it does not grant arbitrary code execution or data exfiltration.

Affected Systems

The vulnerability exists in the vllm project’s vllm engine. All releases prior to version 0.30.0 are affected; the problem is fixed starting with v0.30.0.

Risk and Exploitability

The CVSS score of 4.2 places the issue in the medium risk range. The EPSS score is not available and the vulnerability is not listed in the CISA KEV catalog. Exploitation requires the attacker to have network access to the /score or /rerank endpoint and to know or guess a victim’s X-Request-Id value. A concurrent request bearing the same identifier can overwrite the cache and trigger the vulnerability. Given these prerequisites, the likelihood of exploitation is moderate and the damage is limited to scoring integrity and possible availability problems.

Remediation

No solution or workaround provided in the CVE record.

OpenCVE Recommended Actions

  • Upgrade the vllm engine to version 0.30.0 or later where the cache key is no longer derived from the caller‑controlled header.
  • Configure the application to generate and enforce a server‑side unique identifier for caching instead of accepting client supplied values on the /score and /rerank endpoints.
  • Monitor logs for cache‑miss errors or abnormal scoring results that could indicate abuse of a shared request ID.

Generated by OpenCVE AI on October 7, 2026 at 06:20 UTC.

Tracking

Sign in to view the affected projects.

Advisories

Source ID Title
Github GHSA Github GHSA GHSA-2phq-3phc-84px vLLM: Flash late-interaction scoring caches query embeddings under a caller-controlled request id — cross-request integrity break and induced errors on `/score` and `/rerank`
History
Type Values Removed Values Added
First Time appeared Vllm
Vllm vllm
CPEs cpe:2.3:a:vllm:vllm:*:*:*:*:*:*:*:*
Vendors & Products Vllm
Vllm vllm
Type Values Removed Values Added
Metrics ssvc

{'options': {'Automatable': 'no', 'Exploitation': 'none', 'Technical Impact': 'partial'}, 'version': '2.0.3'}

Type Values Removed Values Added
First Time appeared Vllm-project
Vllm-project vllm
Vendors & Products Vllm-project
Vllm-project vllm
Type Values Removed Values Added
Description vLLM is an inference and serving engine for large language models. Prior to 0.30.0, flash late-interaction scoring at the /score and /rerank endpoints derives each worker's query_key value from the caller-controlled X-Request-Id header. A concurrent request that reuses a victim's identifier can overwrite the cached query embedding so the victim's documents are scored against the attacker's query, and shared use counters can also cause a late-interaction cache-miss error. This issue is fixed in version 0.30.0.
Title vLLM: Flash late-interaction scoring caches query embeddings under a caller-controlled request id — cross-request integrity break and induced errors on `/score` and `/rerank`
Weaknesses CWE-639
References
Metrics cvssV3_1

{'score': 4.2, 'vector': 'CVSS:3.1/AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:L/A:L'}

Weaknesses
  • CWE-639

    Authorization Bypass Through User-Controlled Key

  • CWE-694

    Use of Multiple Resources with Duplicate Identifier

Sources

Status
PUBLISHED
Assigner
GitHub_M
Published
Updated
Oct 6, 2026, 5:55 PM
Reserved
Oct 5, 2026, 7:11 PM
Status
Analyzed
Published
Oct 5, 2026, 11:17 PM
Modified
Oct 8, 2026, 1:28 AM
Severity
Moderate
Public date
Oct 5, 2026, 10:47 PM