CVE-2026-105755

A moderate severity flaw was found in vLLM, the inference and serving engine used by Red Hat AI Inference Server and Red Hat OpenShift AI (RHOAI). When handling flash late-interaction scoring and reranking requests at the /score and /rerank endpoints, vLLM caches each request's query embedding under a key derived from the caller-controlled X-Request-Id header. A remote attacker with API client access who obtains or predicts a victim's request identifier and submits a precisely timed concurrent request reusing it can overwrite the victim's cached query embedding, so the victim's documents are scored against the attacker's query and the scoring results returned to the victim are corrupted. Collisions on shared use counters can also trigger transient late-interaction cache-miss errors, causing denial of service for individual requests. The flaw does not disclose confidential information to the attacker. Red Hat and upstream both rate this issue moderate severity with a CVSS 3.1 score of 4.2 (AV:N/AC:H/PR:L/UI:N/S:U/C:N/I:L/A:L). Red Hat has deferred a fix for the affected Red Hat AI Inference Server and Red Hat OpenShift AI releases because the impact is moderate and the CVSS score is below 7.0; the residual risk is limited to corruption of individual scoring or reranking responses and transient request failures.