What happened
The vLLM Rust gRPC converter fails to enforce the max_logprobs limit when handling CandidateTokens.top_n. A caller can send Generate or GenerateStream requests with oversized or integer-wrapping top_n values that bypass the intended cap and reach downstream processing.
Why it matters
Lets a remote caller exceed configured logprobs limits, which can consume excessive resources and destabilize the serving engine (CVSS 5.9, CWE-20).
Attack vector
Send a Generate/GenerateStream request with oversized or integer-wrapping CandidateTokens.top_n to bypass the max_logprobs cap.
Affected systems
vLLM >= 0.22.0, < 0.24.0 (Rust gRPC converter).
Mitigation
Upgrade to vLLM 0.24.0.