Vulnerability in vllm (CVE-2026-44223)
Summary
Vulnerability in vllm (CVE-2026-44223). Risk of unauthorized operations or information disclosure. Exploitable via extract_hidden_states. Mitigation: upgrade to 0.20.0 or later.
Original description
The description as published by the source.
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,…
Source: OSV.dev (Open Source Vulnerabilities) NVD GitHub Advisory OSV
Auto-generated summary
Assembled by SecPulse from the source data (CVSS, CWE, affected versions) using templates — no AI involved. Check the original description and references below for exact details.
A vulnerability tracked as CVE-2026-44223 has been found in vllm.
Attackers can target a specific entry point like extract_hidden_states over the network to misuse the product.
Risk of unauthorized operations or information disclosure. CVSS score: 6.5/10.
What to do: upgrade vllm to 0.20.0 or later.
If unsure, ask your IT team or search "vllm CVE-2026-44223" on the vendor's site.
CVE-2026-44223 (vllm) — CWE-131 / CVSS v3 6.5
Attack vector: remote (network-reachable) / no user interaction
Attack surface: extract_hidden_states / RuntimeError / repetition_penalty / frequency_penalty
Patched: 0.20.0 — apply immediately
Workaround: avoid using extract_hidden_states as the speculative decoding method on affected versions.
Plan: 1) Audit SBOM/dependencies, 2) Stage→prod upgrade, 3) Add WAF/proxy monitoring on affected endpoints, 4) Hunt IOCs in logs.
Refs: see the GHSA / vendor advisory / patched release linked on this page.
A vulnerability (CWE-131) exists in vllm. Attackers reach the vulnerable code path via extract_hidden_states.
Check the vendor advisory for the affected versions of vllm. Attack surface: extract_hidden_states / RuntimeError / repetition_penalty / frequency_penalty.
Severity: Medium (CVSS 6.5/10). Risk of unauthorized operations or information disclosure.
Update to 0.20.0.
Workaround: avoid using extract_hidden_states as the speculative decoding method on affected versions.
Search webserver/proxy logs for unusual request patterns matching this CVE's known IOCs. Run grep -r 'vllm' . against your dependency files (package-lock.json, requirements.txt, go.sum) to find affected services.
Response actions
Concrete steps and command examples for SOC/SRE teams to execute in order
-
1Identify exposure identify
grep -r 'vllm' . | grep -v node_modulesリポジトリと本番環境の依存ファイル (package-lock.json / requirements.txt / go.sum / Gemfile.lock 等) で
vllmを grep し、稼働しているサービス・バージョンを把握する。 -
2Apply temporary workaround mitigate
avoid using `extract_hidden_states` as the speculative decoding method on affected versions.パッチが適用されるまでの応急処置として、avoid using
extract_hidden_statesas the speculative decoding method on affected versions. を実施。回避策の副作用 (機能低下) を確認した上で。 -
3Apply patch patch
Upgrade vllm to 0.20.0ステージング環境で 0.20.0 に上げて回帰テスト → 本番反映。回帰テストはアプリの主要ハッピーパスを含めること。
-
4Post-deployment verification verify
Confirm patched version is live in productionパッチ適用後、ステージングで PoC または同等の悪用パターンを再現して脆弱性が閉じたことを確認。本番ではアクセスログ・WAF でアラート再発が無いか継続監視。
Affected packages
References
- advisory nvd.nist.gov
- package github.com
- patch github.com
- patch github.com
- web github.com
- web github.com