CVE-2026-114050

What it actually does

Unauthenticated remote attackers can supply arbitrary multimodal UUID values that are used verbatim as cache keys without validation; repeated or colliding UUIDs desynchronize the P0/P1 processor caches, producing a fatal token-placeholder mismatch that terminates the vLLM engine process.

Why it matters to the business

Successful exploitation causes complete loss of the inference service until the process is manually restarted, resulting in downtime for any applications or users relying on the model-serving instance and potential loss of in-flight requests.

How an attack unfolds

An attacker repeatedly submits inference requests containing chosen multimodal UUID values; because the UUID is accepted without consistency checks, the attacker forces cache collisions that desynchronize internal processor state and trigger an unrecoverable token-placeholder mismatch crash.

How to fix it

Upgrade all vLLM deployments to version 0.31.0 or later; after upgrade, restart instances and verify that client-supplied UUIDs are no longer used verbatim for cache identity without validation.

How to confirm the fix worked

Confirm the running binary or container image reports version 0.31.0 or newer; attempt to reproduce the crash with a colliding multimodal UUID and verify the engine no longer terminates.

Typical attack path

1. Identify a reachable vLLM endpoint accepting multimodal inputs
2. Craft and send one or more inference requests containing a chosen or colliding multimodal UUID
3. Observe cache desynchronization and token-placeholder mismatch in server logs
4. Trigger repeated crashes to sustain denial of service

How to detect it

Search application and engine logs for repeated occurrences of token-placeholder mismatch errors, abrupt process terminations, or cache-identity warnings around the time of service unavailability; correlate with inbound requests containing multimodal metadata.

• token-placeholder mismatch
• multimodal uuid
• P0/P1 processor caches
• fatal cache desynchronization

What your security team should do

• Immediately upgrade to vLLM 0.31.0 or newer across all environments
• Add version checks and automated restart policies for inference services
• Monitor crash logs for token-placeholder mismatch messages until patches are applied
• Restrict direct exposure of multimodal endpoints to trusted clients where feasible