Reachable Assertion Affecting vllm package, versions [0.28.0,0.30.0)


Severity

Recommended
0.0
high
0
10

CVSS assessment by Snyk's Security Team. Learn more

Threat Intelligence

EPSS
0.37% (32nd percentile)

Do your applications use this vulnerable package?

In a few clicks we can analyze your entire application and see what components are vulnerable in your application, and suggest you quick fixes.

Test your applications
  • Snyk IDSNYK-PYTHON-VLLM-22556259
  • published15 Nov 2026
  • disclosed13 Nov 2026
  • creditwildoranges

Introduced: 13 Nov 2026

NewCVE-2026-126398  (opens in a new tab)
CWE-617  (opens in a new tab)

How to fix?

Upgrade vllm to version 0.30.0 or higher.

Overview

vllm is an A high-throughput and memory-efficient inference and serving engine for LLMs

Affected versions of this package are vulnerable to Reachable Assertion via a reachable assert in GPUModelRunner._init_xdrope_positions that crashes the EngineCore and API server process when an embeddings-only request is submitted to an XD-RoPE model. The assertion requires req_state.prompt_token_ids to be non-None, but embeddings-only /v1/completions requests legitimately reach the worker with prompt_token_ids = None, causing an unhandled AssertionError that terminates the EngineCore. Because the EngineCore dies, the API server shuts down and all concurrent clients across all tenants are disconnected.

Note: This is only exploitable when the server is running an XD-RoPE model (rope_scaling.type == "xdrope") with the --enable-prompt-embeds flag enabled on the V1 (MRV1) GPU model runner.

References

CVSS Base Scores

version 4.0
version 3.1