CVE-2026-116172: vLLM: gRPC top_n validation flaw crashes EngineCore

A flaw in the Rust gRPC converter of vLLM 0.22.0 up to (but not including) 0.24.0 fails to validate CandidateTokens.top_n, so a remote caller can bypass max_logprobs limits and send oversized or integer-wrapping values that make torch.topk fail and terminate EngineCore. The impact is availability only (CVSS 5.9, no confidentiality or integrity impact), and there is no known public exploit, no Nuclei template, and no CISA KEV listing; EPSS data is not yet available. The risk is higher than the medium score suggests for any organization that exposes the vLLM gRPC endpoint to untrusted networks, because one crafted Generate or GenerateStream request can take down the whole inference engine rather than a single request. Upgrade to vLLM 0.24.0 or later, restrict the gRPC port to trusted clients behind authentication, and alert on unexpected EngineCore restarts.

No confirmed active exploitation of CVE-2026-116172 has been reported, but organizations should still patch proactively.

1) Upgrade vLLM to 0.24.0 or later. 2) Until patched, restrict network access to the gRPC port to trusted clients using network policies, a service mesh, or an authenticated API gateway. 3) Add a gateway or proxy rule that rejects Generate and GenerateStream requests with unusually large top_n values or values that exceed max_logprobs. 4) Monitor for EngineCore terminations, torch.topk errors, and unexpected pod or process restarts, and alert on repeated crash loops. 5) Configure rate limiting and health checks so failed instances are removed from load balancers and restarts are throttled. 6) Review the GitHub advisory GHSA-rwfw-xvg3-3937 for further guidance.

This vulnerability affects the following AI/ML architecture patterns: model serving, LLM inference endpoints, gRPC-based inference APIs, RAG pipelines, agent backends.

CVE-2026-116172 has a CVSS v3.1 base score of 7.5 (HIGH). The EPSS exploitation probability is 0.34%.