What happened
The vLLM Rust gRPC Generate/GenerateStream service forwards unchecked prompt.token_ids to EngineCore without validating them against the model vocabulary. Requests with out-of-vocabulary prompt.token_ids can trigger a fatal embedding lookup and crash the engine.
Why it matters
A single malformed request can crash the shared engine, denying service to other clients (CVSS 5.9, CWE-20).
Attack vector
Send a request with out-of-vocabulary prompt.token_ids to trigger a fatal embedding lookup and crash the engine.
Affected systems
vLLM >= 0.22.0, < 0.24.0 (Rust gRPC Generate/GenerateStream).
Mitigation
Upgrade to vLLM 0.24.0.