vLLM flaw reaches version 0.29.0

vLLM has a newly published medium-severity CVE affecting versions through 0.29.0: CVE-2026-93841 can corrupt sampler state across concurrent requests during multimodal inference. VulnCheck’s advisory rates it CVSS 6.3, while the upstream fix proposal is still open; the public record does not yet name a released version containing the fix. Operators should inventory vLLM 0.29.0-and-earlier multimodal deployments and pin a vendor-confirmed patched release before treating the serving stack as closed—especially where one bad request can contaminate neighboring work. That makes Temporal’s storage gap for agent histories part of the same reliability ledger: inference correctness and state isolation are both production concerns.