Improper Input Validation in vLLM via top_n Parameter

Executive Summary

This vulnerability affects vLLM versions 0.22.0 to 0.23.x. It involves improper input validation in the Rust gRPC converter where the CandidateTokens.top_n parameter bypasses max_logprobs limits. Attackers can send requests with oversized or integer-wrapping top_n values, causing torch.topk failures that terminate the EngineCore, leading to denial of service.

Useful 0 Not Useful 0

Detection Guidance

Monitor vLLM logs for EngineCore termination errors or torch.topk failures during Generate or GenerateStream requests. Check for requests with unusually large CandidateTokens.top_n values (e.g., 1,000,000 or 2,147,483,648) in gRPC traffic. Use network monitoring tools to inspect gRPC payloads for oversized top_n parameters.

Useful 0 Not Useful 0

Impact Analysis

The primary impact is denial of service. Attackers can send malicious requests to crash the vLLM service by exploiting the unbounded top_n parameter, which causes EngineCore to terminate unexpectedly. This disrupts service availability for legitimate users.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM to version 0.24.0 or later to patch the vulnerability. If upgrading is not immediately possible, restrict network access to vLLM services or implement rate limiting to reduce exposure. Validate and sanitize all gRPC requests to ensure top_n values are within safe bounds.

Useful 0 Not Useful 0