ZeroDay Findings
Sign inSign up

Findings

CVE-2026-105754

vLLM: Scale-out disaggregated multimodal transport trusts caller-supplied features

Published Oct 5, 2026 · corrected/updated Oct 15, 2026 · evidence confidence high

## Affected - **Ecosystem / package:** pip / - **Affected versions:** vLLM ≤ 0.25.1 (confirmed on 0.25.1, commit [ ](https://github.com/vllm-project/vllm/tree/752a3a504485790a2e8491cacbb35c137339ad34)). The lower bound predates 0.25.1; maintainers can confirm how far back the scale-out transport path reaches. ## Summary vLLM's disaggregated **scale-out** transport splits a multimodal request into a trusted render step ( ) and a separate generate step ( ). The generate route decodes a caller-supplied object — serialized encoder tensors ( ), multimodal hashes ( ), placeholder ranges ( ), and the internal field-processor selection — and forwards it into the engine **as if it had come from the trusted renderer**, with no rebinding to (or validation against) the active model's renderer contract. Because the two routes are ordinary auth-guarded HTTP endpoints (the prefix is registered by default on generate-capable servers), any authenticated caller can submit an otherwise-valid render body with a single forged field. Depending on which field is forged, this produces: - an **engine-fatal crash** of the shared EngineCore process (denial of service), reproduced as a CUDA illegal-memory-access, a post-admission rank-mismatch , and a hard — three independent forged fields (sites 1, 2, 3); - **silent cross-request encoder-cache poisoning / disclosure** of shared encoder state when the cache-key hash is not bound to the payload (site 4); - **transport-level integrity loss** when sparse placeholder masks are dropped during render-to-generate replay (site 5). All five share one root cause and one fix shape: the reconstructed multimodal state on the scale-out path is trusted without being rebound to, and validated against, the active model's renderer output before it reaches the engine. These sites are distinct from prior multimodal hardening. Site 1 survives [GHSA-wv77-2vpf-vmmg](https://github.com/vllm-project/vllm/security/advisories/GHSA-wv77-2vpf-vmmg) (that fix validates ful

Grounded briefings

Catalog interpretations for three audiences. Each brief cites ingested metadata and the original advisories below. They are not independent research and they do not invent exploit steps.

Executive

CVE-2026-105754 is a medium public vulnerability record for this product. It is not currently listed in the CISA Known Exploited catalog. Check the linked vendor or CERT advisories for patch status. This summary is a catalog interpretation of the ingested metadata and cites the original advisories below.

System administrator

Identify whether this product is in your estate, then compare installed versions with the affected ranges. Affected packages or ranges on record: vllm (0 → 0.30.0). Check the linked vendor or CERT advisories for patch status. Prefer the vendor or distribution advisory linked in Sources before changing production systems. This is a catalog interpretation, not a runbook.

Security analyst

Signals on file: CVSS 6.5, EPSS 0.3%. It is not currently listed in the CISA Known Exploited catalog. Weakness classes: CWE-1284, CWE-20, CWE-617, CWE-639. Sources contributing to this record: cve_org, epss, github, osv. Public exploit references, when present, are outbound links only and are not proof of active exploitation. This is a catalog interpretation of stored metadata.

Exploitation and remediation

  • Exploitation evidence

    Not listed as known exploited in CISA KEV. Not in CISA KEV or VulnCheck KEV.

  • Patch and remediation

    No required-action text is stored yet. Use the vendor or CERT links in Sources for patch availability.

  • Safe detection guidance

    Compare installed versions with the affected ranges below and verify the advisory from the original publisher. This site stores metadata and outbound links only; it does not ship exploit code or proof-of-concept payloads.

Code

The project this record applies to. Exploit listings stay in References.

Timeline

Detection, updates, and resolution for this record.

  1. Oct 5, 2026 · GitHub Advisory Database

    Detected

    ## Affected - **Ecosystem / package:** pip / - **Affected versions:** vLLM ≤ 0.25.1 (confirmed on 0.25.1, commit [ ](https://github.com/vllm-project/vllm/tree/752a3a504485790a2e8491cacbb35c137339ad34)). The lower bound predates 0.25.1; maintainers can confirm how far back the scale-out transport path reaches. ## Summary vLLM's disaggregated **scale-out** transport splits a multimodal request into

  2. Oct 15, 2026 · GitHub Advisory Database

    Fix available

    vllm fixed in 0.30.0

  3. Oct 15, 2026 · GitHub Advisory Database

    Fix available

    vllm fixed in 0.30.0

  4. Oct 15, 2026 · GitHub Advisory Database

    Updated

    ## Affected - **Ecosystem / package:** pip / - **Affected versions:** vLLM ≤ 0.25.1 (confirmed on 0.25.1, commit [ ](https://github.com/vllm-project/vllm/tree/752a3a504485790a2e8491cacbb35c137339ad34)). The lower bound predates 0.25.1; maintainers can confirm how far back the scale-out transport path reaches. ## Summary vLLM's disaggregated **scale-out** transport splits a multimodal request into

Affected

  • PyPI vllm0 → 0.30.0

Sources

References