CVE-2026-44223
MediumWhat is CVE-2026-44223?
vLLM is an inference and serving engine for large language models (LLMs). From to before 0.20.0, 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. This vulnerability is fixed in 0.20.0.
CVSS Vector Breakdown
CVSS 3.1Exploitable remotely over the network without requiring access to the local system.
No special conditions required. Exploit can be performed reliably and repeatedly.
Basic user account required (e.g. standard login without admin rights).
No user interaction required. Exploit runs without any victim action.
Exploiting the vulnerability only affects the vulnerable component itself.
No confidentiality impact. Sensitive data is not exposed.
No integrity impact. Data cannot be modified.
Complete denial of service. The component becomes fully unavailable.
FAQ
CVE-2026-44223 has a CVSS score of 6.5/10, rated as Medium. Review the affected products list and apply vendor patches.
CVSS (Common Vulnerability Scoring System) rates vulnerability severity from 0.0 to 10.0. CVE-2026-44223 scores 6.5/10 (Medium). Scores 9.0–10.0 are Critical, 7.0–8.9 are High, 4.0–6.9 are Medium, and below 4.0 are Low.
The list of devices confirmed to be affected by CVE-2026-44223 is shown in the "Affected Devices" section above. Check your firmware version against the vendor security advisory and apply the latest patch.
Apply the latest firmware or software update from the vendor. Check the References section above for official advisories and patch notes. If no patch is available, consider disabling the affected feature or isolating the device from untrusted networks.