CVE-2026-105755

NVD: vLLM is an inference and serving engine for large language models. NVD: 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. NVD: 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.

Source-published summary

NVD: vLLM is an inference and serving engine for large language models. NVD: 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. NVD: 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.

Possible impact

This medium severity issue needs human triage to confirm exposure, affected versions, and vendor guidance for defensive exposure review.

Affected context

vendor/product: vllm / vllm

Remediation / advisory

Patch confirmed by source text; fixed version context: 0.30.0..

Why it matters

This medium severity issue needs human triage to confirm exposure, affected versions, and vendor guidance for defensive exposure review.; CVSS 4.2 (MEDIUM); EPSS percentile 9; not listed in KEV; Patch confirmed by source text; fixed version context: 0.30.0; sources: NVD, OSV, Vendor Advisory.

What to verify

Confirm affected product/version, vendor advisory, patch or mitigation, and exposure.

Exposure hint

exposure unknown

Impact tags

Urgency reasons

Source-derived note

Summary derived from NVD / OSV / Vendor Advisory description; unsafe procedural detail is not shown.

Redaction metadata

source summary used

True

fallback summary used

False

unsafe procedural detail present

false

raw source displayed

false

public summary redacted

true

Remediation handoff

Public-safe static handoff for human/Codex remediation planning. Scan, patch, external execution, and auto remediation are disabled.

Safety note

This radar shows source-published defensive context only. Exploit procedures, exploit strings, scanner commands, and auto-remediation are not provided.

Official references

• https://nvd.nist.gov/vuln/detail/CVE-2026-105755
• https://osv.dev/vulnerability/CVE-2026-105755
• https://github.com/vllm-project/vllm/commit/ee17d0d869203ef9a35ad73358a4987bba14b1fc
• https://github.com/vllm-project/vllm/pull/51445
• https://github.com/vllm-project/vllm/releases/tag/v0.30.0
• https://github.com/vllm-project/vllm/security/advisories/GHSA-2phq-3phc-84px