What happened
The vLLM disaggregated /inference/v1/generate endpoint accepts inconsistent multimodal metadata: a client can replay a valid multimodal item with repeated or overlapping mm_placeholder values, which is not properly validated and can crash the engine.
Why it matters
A remote caller can crash the serving engine in a disaggregated (P/D) deployment, denying service to other requests (CVSS 5.9, CWE-20).
Attack vector
Replay a valid multimodal item with repeated/overlapping mm_placeholder metadata to the /inference/v1/generate endpoint.
Affected systems
vLLM >= 0.20.0, < 0.30.0 (disaggregated multimodal path).
Mitigation
Upgrade to vLLM 0.30.0.