CVE-2026-56340: vLLM introduced enhanced protection for CVE-2025-62164
vLLM 0.10.2 up to but not including 0.13.0 does not validate sparse tensors in embeddings it receives, so a crafted request with negative or out-of-bounds indices can crash the server. The earlier fix for CVE-2025-62164 only turned prompt embeds off by default. Servers that enable the feature stay exposed until 0.13.0.
Published Jan 8, 2026 · Updated Sep 11, 2026 · Analysis Nov 1, 2026
What an attacker can do
PyTorch disables sparse tensor invariant checks by default for speed, so a malformed sparse tensor passes in unchecked. vLLM accepts embeddings from clients, and the earlier fix for CVE-2025-62164 (pull request 27204) added a flag that disables prompt embeds by default. It did not add validation, so once an operator turns the feature on, the original problem is back.
An authenticated client sends embeddings with negative or out-of-range indices. The result is a crash or resource exhaustion, and the VulnCheck write-up also warns of out-of-bounds memory writes that could corrupt memory. The CWEs include out-of-bounds write and deserialization of untrusted data. The CVSS vector is network, low privileges, no user interaction, high impact on confidentiality, integrity and availability. In practice the sure outcome is denial of service for every user of the server; code execution is a risk the advisory flags but does not demonstrate. Servers that never enabled prompt embeds are not exposed.
Who is affected
The vllm pip package from 0.10.2 up to 0.13.0 (exclusive), when the prompt-embeds feature is enabled. Check the version, then check the launch flags and config of each server for prompt embeds:
pip show vllmvllm
- Ecosystem
- PyPI
- Affected
- >= 0.10.2, < 0.13.0
- Fixed in
- 0.13.0
- Check
pip show vllm
How to fix it
Upgrade to vLLM 0.13.0 or later:
pip install --upgrade "vllm>=0.13.0"The change (pull request 30649) adds sparse tensor validation so that indices are checked to be valid, non-negative and within bounds. The sources give no other workaround. Until you upgrade, leave prompt embeds disabled, which is the default after the earlier fix, and limit who can reach the API.
How to tell if you were hit
Check for engine crashes or restarts, and for unusual memory growth, in servers that have prompt embeds enabled. Review request logs for embedding payloads from one client just before a crash. Search your launch scripts and configs for the prompt-embeds option to find servers that are exposed in the first place.
Original advisory
From the GitHub Advisory Database (CC BY 4.0).
Summary
The fix here for CVE-2025-62164 is not sufficient. The fix only disables prompt embeds by default rather than addressing the root cause, so the DoS vulnerability remains when the feature is enabled.
Details
vLLM's pending change attempts to fix the root cause, which is the missing sparse tensor validation. PyTorch (~v2.0) disables sparse tensor validation (specifically, sparse tensor invariants checks) by default for performance reasons. vLLM is adding the sparse tensor validation to ensure indices are valid, non-negative, and within bounds. These checks help catch malformed tensors.
PoC
NA
Impact
Current fix only added a flag to disable/enable prompt embeds, so by default, prompt embeds feature is disabled in vLLM, which stops DoS attacks through the embeddings. However, It doesn’t address the problem when the flag is enabled and there is still potential for DoS attacks.
Changes
- https://github.com/vllm-project/vllm/pull/30649
Questions
- Is my server affected if I never enabled prompt embeds?
The advisory says the earlier fix disabled the feature by default and that the denial of service remains only when it is enabled. Check your own flags to be sure.
- How does this relate to CVE-2025-62164?
Pull request 27204 fixed that CVE by turning off prompt embeds by default. This advisory covers the missing root cause fix, the sparse tensor validation.
- Which version fixes it?
vLLM 0.13.0 fixes it. The affected range is 0.10.2 up to but not including 0.13.0, and the fix adds sparse tensor validation for incoming embeddings.
References
Read for this analysis:
- https://github.com/vllm-project/vllm/security/advisories/GHSA-mcmc-2m55-j8jj
- https://www.vulncheck.com/advisories/vllm-denial-of-service-via-unvalidated-multimodal-embeddings
- https://github.com/vllm-project/vllm/pull/30649
- https://access.redhat.com/security/cve/CVE-2026-56340
- https://bugzilla.redhat.com/show_bug.cgi?id=2491060
- https://github.com/pypa/advisory-database/tree/main/vulns/vllm/PYSEC-2026-250.yaml
- https://security.access.redhat.com/data/csaf/v2/vex/2026/cve-2026-56340.json
- https://github.com/advisories/GHSA-mcmc-2m55-j8jj
How this page is made: the data panel comes from the GitHub Advisory Database, CISA’s KEV catalog, the NVD and FIRST EPSS, and updates twice a day. The analysis is written by an AI model from the sources listed above and checked by automated rules before it is published. Check versions against the vendor's advisory before you act. Methodology
More vLLM vulnerabilities
All vLLM entriesCVE-2026-115911
vLLM: DoS via unbounded chat_template Jinja2 rendering exhausts the renderer thread pool
medium6.5Oct 28, 2026CVE-2026-115912
vLLM: Public multimodal SHM handle forgery reaches worker-side `pickle.loads()` with insecure serialization disabled
high7.5Oct 28, 2026CVE-2026-115913
vLLM: Phi-4 multimodal audio encoder materializes and retains full attention matrices: a single max-length audio OOM-crashes the vLLM engine
medium6.5Oct 28, 2026CVE-2026-115914
vLLM DeepSeek-OCR image processing vulnerable to denial of service
medium6.5Oct 28, 2026CVE-2026-115915
vLLM: Untrusted EPS inputs can trigger denial of service or Ghostscript RCE
high8.1Oct 28, 2026CVE-2026-115916
vLLM: Out-of-vocabulary prompt token id passes input validation and triggers a CUDA device-side assert (engine crash / DoS)
medium6.5Oct 28, 2026
