Your RSA-2048 keys break in 2030. Find every one of them before attackers do.
🐍
🐍 PyPI
Not in CISA KEV
MEDIUM severity

CVE-2026-44223 — vllm

MEDIUMFix: vllm-project/vllm#38610

CVE-2026-44223 is a medium-severity (CVSS 6.5) CWE-131 vulnerability in vllm. A fix is available for vllm — see the affected versions and patch details below.

vLLM: extract_hidden_states speculative decoding crashes server on any request with penalty parameters

Also known asGHSA-83vm-p52w-f9pwPYSEC-2026-145
Published
May 12, 2026
Updated
Sep 4, 2026
Affected
1 pkg
Patched
1 / 1
Exploits
None indexed
Exploitation data as of Sep 24, 2026 · OSV.dev, NVD, FIRST.org (EPSS)

Exploitation Status

Proof-of-concept exploit code exists

  • CISA’s SSVC triage found public proof-of-concept exploit code for this CVE, though no confirmed active exploitation.

Exploitation and automatability from CISA’s SSVC triage for CVE-2026-44223.

EPSS Exploitation Probability

via FIRST.org ↗
0.4%probability of exploitation in next 30 days
Lower Risk0.00%
Lower risk than most CVEs34th percentile — riskier than 34% of all scored CVEsHighest risk

EPSS (Exploit Prediction Scoring System) is a daily probability model maintained by FIRST.org. It estimates the likelihood a CVE will be exploited in production environments within the next 30 days, derived from real-world threat intelligence signals.

How urgent is this, really

CVE-2026-44223 plotted by exploitation likelihood (EPSS) against impact (CVSS). The shaded corner — EPSS 50%+ and CVSS 7.0+ — is where this CVE doesn't sit, though severity or exploitability alone can still warrant action.

Where this sits among everything scored

Of 378,567 CVEs with a current EPSS score, this one falls in the < 10% band (highlighted). Real counts from FIRST.org, not a sample — log-scaled since the landscape is heavily right-skewed.

Real-World Exposure

1 pkg affected
🐍vllm

Real-time download stats are indexed for npm and PyPI packages. This vulnerability affects PyPI packages — download data is not available via public APIs for these ecosystems.

Description

Summary

The extract_hidden_states speculative decoding proposer in vLLM returns a tensor with an incorrect shape after the first decode step, causing a RuntimeError that crashes the EngineCore process. The crash is triggered when any request in the batch uses sampling penalty parameters (repetition_penalty, frequency_penalty, or presence_penalty).

A single request with a penalty parameter (e.g., "repetition_penalty": 1.1) is sufficient to crash the server. The crash is deterministic and immediate — no concurrency, race condition, or special workload is required.

Details

In vLLM v0.17.0, the extract_hidden_states proposer's propose() method returned sampled_token_ids.unsqueeze(-1), producing a tensor of shape (batch_size, 1).

In PR #37013 (first released in v0.18.0), the KV connector interface was refactored out of propose(). The return type changed from tuple[Tensor, KVConnectorOutput | None] to Tensor, and the .unsqueeze(-1) call was removed along with the KV connector output:

# Before (v0.17.0):
return sampled_token_ids.unsqueeze(-1), kv_connector_output  # shape (batch_size, 1)

# After (v0.18.0+):
return sampled_token_ids  # shape (batch_size, 2) after first decode step

The refactor missed that sampled_token_ids changed semantics between the first and subsequent decode steps. After the first decode step, the rejection sampler allocates its output as (batch_size, max_spec_len + 1). With num_speculative_tokens=1, this produces shape (batch_size, 2) instead of the expected (batch_size, 1), causing a broadcast shape mismatch during penalty application.

Impact

Any vLLM deployment between v0.18.0 and v0.19.1 (inclusive) configured with extract_hidden_states speculative decoding is affected. A single API request containing any penalty parameter immediately and permanently crashes the EngineCore process, resulting in complete loss of service availability.

Patches

Fixed in PR #38610, first included in vLLM v0.20.0. The fix slices the return value to sampled_token_ids[:, :1], ensuring the correct (batch_size, 1) shape regardless of the rejection sampler's output dimensions.

Workarounds

  • Upgrade to vLLM v0.20.0 or later.
  • If upgrading is not possible, avoid using extract_hidden_states as the speculative decoding method on affected versions.
  • Alternatively, reject or strip penalty parameters (repetition_penalty, frequency_penalty, presence_penalty) from incoming requests at an API gateway before they reach vLLM.

Affected Packages

1 total 1 fixed
EcosystemPackageVulnerable rangeFix
🐍PyPIvllm≥ 0.18.0&&< 0.20.00.20.0pip install --upgrade 'vllm==0.20.0'

Detection & mitigation playbook

Open-source dependency
  1. Detect

    Scan your dependency tree (package-lock.json, pnpm-lock.yaml, requirements.txt, go.sum, etc.) for vllm, including transitive dependencies — a direct dependency you never call can still pull in a vulnerable version.

  2. Fix

    Update vllm to 0.20.0 or later, then make sure no transitive (indirect) dependency still pins the vulnerable range — O3 confirms CVE-2026-44223 is resolved across your whole dependency graph.

  3. Workarounds

    If you can't upgrade right away: gate or disable the affected feature, validate untrusted input at the boundary, and avoid passing attacker-controlled data into the vulnerable path. O3's runtime protection blocks exploitation in production as an interim safeguard until the upgrade lands.

  4. How O3 protects you

    O3 Security's impact-aware SCA analyses which vulnerable code paths your application actually calls, so a match like CVE-2026-44223 can be triaged on real exposure rather than presence alone.

Tailored to CVE-2026-44223. Runtime protection reduces exposure until a permanent patch is applied and verified — it complements patching, it doesn't replace it.

Fixing This On Your OS

If you run this on a Linux distribution, patch through your package manager against the distro's own security advisory below — it tracks the exact backported fix for your release, which can ship on a different timeline (and sometimes a different severity) than the upstream project.

Red HatImportant

This Important denial of service flaw in vLLM, as used in Red Hat AI Inference Server and Red Hat OpenShift AI, allows a remote attacker to crash the EngineCore process. By sending a request with specific sampling penalty parameters, an attacker can trigger an incorrect tensor shape, leading to a service disruption…

ProductFixed inAdvisory
Red Hat AI Inference Server 3.4rhaii/vllm-cpu-rhel9:1787151769RHSA-2026:57380
Red Hat AI Inference Server 3.4rhaii/vllm-spyre-rhel9:1787151840RHSA-2026:57387
Red Hat AI Inference Server 3.4rhaii/vllm-cuda-rhel9:1787151771RHSA-2026:57389
Red Hat AI Inference Server 3.4rhaii/vllm-rocm-rhel9:1787151774RHSA-2026:57390
Red Hat AI Inference Server 3.4rhaii/vllm-spyre-rhel9:1789681201RHSA-2026:69464
Red Hat AI Inference Server 3.4rhaii/vllm-cpu-rhel9:1789681128RHSA-2026:69466
Red Hat AI Inference Server 3.4rhaii/vllm-cuda-rhel9:1789681126RHSA-2026:69467
Red Hat AI Inference Server 3.4rhaii/vllm-rocm-rhel9:1789681126RHSA-2026:69469

Frequently Asked Questions

### Summary The `extract_hidden_states` speculative decoding proposer in vLLM returns a tensor with an incorrect shape after the first decode step, causing a `RuntimeError` that crashes the EngineCore process. The crash is triggered when any request in the batch uses sampling penalty parameters (`repetition_penalty`, `frequency_penalty`, or `presence_penalty`). A single request with a penalty parameter (e.g., `"repetition_penalty": 1.1`) is sufficient to crash the server. The crash is deterministic and immediate — no concurrency, race condition, or special workload is required. ### Details
O3 Security · Impact-Aware SCA

Is CVE-2026-44223 in your dependencies?

O3 Security finds CVE-2026-44223 across PyPI dependencies, including transitive ones, and its impact-aware SCA ranks findings by whether your code actually calls the vulnerable path.

CVE-2026-44223: vllm (Medium 6.5) | O3 Security