CVE-2026-105756

A moderate severity denial-of-service vulnerability was found in vLLM, as shipped in Red Hat AI Inference Server and Red Hat OpenShift AI (RHOAI). vLLM's OpenAI-compatible API endpoints accept a client-supplied cache_salt parameter that is validated only as a non-empty string, without character or length restrictions. On deployments where the LMCache-MP KV connector is enabled, the value is forwarded unvalidated into the scheduler's cache lookup, where the downstream LMCache library raises a ValueError for salts that contain forbidden characters (for example '@', '/', backslash, or NUL) or exceed 128 characters. Because this exception is not handled at the request level, the EngineCore process treats it as fatal and terminates, disrupting service for all concurrent users. An authenticated remote attacker with access to the inference API can trigger this crash with a single crafted request. The flaw affects availability only, with no impact on confidentiality or integrity (CVSS 3.1 score of 6.5, AV:N/AC:L/PR:L/UI:N/S:U/C:N/I:N/A:H), which matches the upstream severity rating. Because the impact is moderate and the CVSS score is below 7.0, the fix for the affected Red Hat AI Inference Server and Red Hat OpenShift AI components has been deferred. The residual risk is limited to a denial of service against a single vLLM serving instance by an already-authenticated API user, and only on deployments that use the LMCache-MP connector.