CVE-2026-93436

Name of the Vulnerable Software and Affected Versions vLLM versions prior to 0.29.1

Description In prefill/decode disaggregated deployments, the software fails to properly clean up decode-side metadata for rejected inference requests. A remote attacker can submit multiple requests using the max tokens parameter set to 0, leading to unbounded memory growth on the decode worker. This results in memory exhaustion and a denial of service (DoS) that persists until the worker restarts. Non-disaggregated deployments are not affected.

Recommendations Update vLLM to a version later than 0.29.0. Avoid using the max tokens parameter set to 0 in requests to the affected API endpoint until the update is applied.