AI Stack Alert High CVE-2026-126398 For AI Eng, SOC / IR, AppSec AI-assisted
CVE-2026-126398: vLLM 0.28.0 to before 0.30.0 reachable assertion lets remote attackers crash engine
vLLM from 0.28.0 before 0.30.0 contains a reachable assertion in GPUModelRunner._init_xdrope_positions. A remote attacker can crash the engine by sending a /v1/completions request containing only prompt_embeds to XD-RoPE models run with --enable-prompt-embeds. The crash terminates the EngineCore and API server for all clients. The CVSS base score is 8.2.
Why it matters
Any exposed vLLM deployment serving XD-RoPE models with prompt embeddings enabled can be taken offline by one unauthenticated request. This affects availability for all users of the service.
Do this
- Upgrade vLLM to version 0.30.0 or later.
- Until patched, disable --enable-prompt-embeds on servers hosting XD-RoPE models.
- Restrict access to /v1/completions to trusted clients at the gateway or reverse proxy.
- Monitor logs for EngineCore termination and API server restarts to detect crashes.
Exploit Clock
No signals yet
- No exploitation signals yet
Signals
- NoCISA KEV
- 8.2CVSS
- NoPublic PoC
- 36/100Priority
- 80%Confidence
Evidence
“vLLM from 0.28.0 before 0.30.0 contains a reachable assertion vulnerability in GPUModelRunner._init_xdrope_positions that allows remote attackers to crash the engine”
“Attackers can send a /v1/completions request containing only prompt_embeds to XD-RoPE models with --enable-prompt-embeds, terminating the EngineCore and API server for all clients.”
Sources
Affected / related
vLLM, vLLM
Published 2026-11-13 10:17 UTC · updated 2026-11-13 12:35 UTC · AI-assisted summary, checked against quoted sources · How we score