AI Stack Alert High CVE-2026-116176 For AI Eng, AppSec, SOC / IR AI-assisted
vLLM before 0.24.0 SSRF via X-Request-Id header in P2pNcclConnector (CVE-2026-116176)
vLLM versions from 0.9.2 before 0.24.0 contain a server-side request forgery flaw. API callers can force outbound connections by embedding a decode address in the X-Request-Id header. This can make a P2pNcclConnector prefill worker connect over ZMQ to an attacker-controlled host.
Why it matters
Any API caller who can reach the affected service can trigger outbound connections from prefill workers. The source reports this can leak the NCCL unique_id and KV-cache tensors, or block the worker. The CVSS base score is 8.8.
Do this
- Upgrade vLLM to 0.24.0 or later on all prefill and decode nodes.
- Until patched, check whether any deployment uses the P2pNcclConnector for disaggregated prefill, and restrict API access to trusted clients.
- At the API gateway or reverse proxy, strip or reject X-Request-Id values that contain embedded addresses or unexpected formats.
- Restrict outbound network egress from prefill workers to known decode peers and block unexpected ZMQ connections.
- Hunt for prefill worker outbound connections to hosts outside the expected cluster, and for unexpected worker stalls.
Exploit Clock
Exploitation unlikely (for now)
- EPSS 0.32%
Signals
- NoCISA KEV
- 8.8CVSS
- 0.32%EPSS
- NoPublic PoC
- 34/100Priority
- 70%Confidence
Evidence
“vLLM from 0.9.2 before 0.24.0 contains a server-side request forgery vulnerability that allows API callers to force outbound connections by embedding a decode address in the X-Request-Id header.”
“Attackers can make a P2pNcclConnector prefill worker ZMQ-connect to an attacker host, leaking the NCCL unique_id and KV-cache tensors or blocking the worker.”
Sources
Affected / related
vLLM, P2pNcclConnector
Published 2026-10-25 14:16 UTC · updated 2026-11-08 02:49 UTC · AI-assisted summary, checked against quoted sources · How we score