AI Stack Alert High CVE-2026-116173 For SOC / IR, AI Eng, AppSec, CISO AI-assisted
vLLM 0.22.0 to <0.24.0: Rust gRPC frontend DoS crashes EngineCore via logprobs
vLLM versions from 0.22.0 before 0.24.0 have a denial of service flaw in the Rust gRPC frontend (CVE-2026-116173). A remote attacker can bypass logprob_token_ids limits by sending Generate or GenerateStream requests with output_logprobs enabled and more than 128 output_candidates.token_ids. This triggers a fatal sampler exception that terminates EngineCore. The source reports a CVSS base score of 8.2.
Why it matters
Any deployment exposing the vLLM gRPC frontend to untrusted clients can lose its inference engine through a single crafted request type, interrupting service for all users. The source does not describe data exposure or code execution.
Do this
- Inventory vLLM installations and identify any running version from 0.22.0 up to but not including 0.24.0.
- Upgrade affected vLLM deployments to 0.24.0 or later.
- Until patched, restrict network access to the gRPC frontend to trusted clients and enforce authentication and rate limits.
- Where feasible, reject Generate and GenerateStream requests that combine output_logprobs with more than 128 output_candidates.token_ids at a gateway or proxy.
- Alert on unexpected EngineCore exits or process restarts in vLLM logs and correlate them with gRPC request sources.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.34%
Signals
- NoCISA KEV
- 8.2CVSS
- 0.34%EPSS
- NoPublic PoC
- 34/100Priority
- 70%Confidence
Evidence
“vLLM from 0.22.0 before 0.24.0 contains a denial of service vulnerability that allows remote attackers to terminate EngineCore”
“Attackers can send Generate or GenerateStream requests with output_logprobs enabled and more than 128 output_candidates.token_ids, triggering a fatal sampler exception that crashes the engine.”
Sources
Affected / related
vLLM, EngineCore, Rust gRPC frontend
Published 2026-10-25 14:16 UTC · updated 2026-11-08 02:50 UTC · AI-assisted summary, checked against quoted sources · How we score