vLLM inference server — multiple medium-severity resource-exhaustion and info-disclosure flaws (5 CVEs)

NVD published five Medium-severity vLLM CVEs on 2026-08-13 covering unbounded prompt inputs causing DoS, a concurrency race in prompt-embedding validation, ReDoS via unvalidated regex parameters, an integer-overflow kernel bug allowing cross-request data leakage in a batched multi-tenant inference request, and traceback information disclosure in error handling.

vLLM is a core piece of AI inference infrastructure at many organizations; the cross-tenant data leakage bug (CVE-2026-73558) in particular is significant because it can let one user's batched inference request read another user's partial input/output on a shared serving instance — a serious multi-tenancy isolation failure in production LLM serving.

CVE-2026-73559 (CVSS 6.5): unbounded prompt list in /v1/completions CompletionRequest.prompt causes resource exhaustion; CVE-2026-73557 (CVSS 6.3): torch.sparse global-state race in safe_load_prompt_embeds; CVE-2026-73556 (CVSS 5.3): unbounded regex in structured-output backend enables ReDoS; CVE-2026-73558 (CVSS 5.3): integer overflow in activation_kernels.cu act_and_mul_kernel can leak another batched user's partial output/input across tenants; CVE-2026-73555 (CVSS 5.3): unsanitized traceback strings leak internal details via validation_exception_handler

vLLM 0.19.0–0.26.0 (varies by CVE, up to 0.27.0 for CVE-2026-73558)

Upgrade vLLM to the fixed versions referenced in each GitHub commit; monitor vllm-project security advisories