Improper Input Validation in vLLM Engine

Executive Summary

vLLM versions from 0.20.1 to 0.29.0 have an improper input validation flaw. Attackers can exploit this by sending requests with out-of-vocabulary stop_token_ids, causing the engine to crash or write outside the V2 runner's logits buffer. This is triggered via OpenAI chat, completion, generate, or batch requests with invalid IDs.

Useful 0 Not Useful 0

Detection Guidance

Detecting this vulnerability requires monitoring for requests containing out-of-vocabulary stop_token_ids. Check vLLM engine logs for crashes or buffer overflow errors. Inspect network traffic for malformed OpenAI API requests targeting chat, completion, generate, or batch endpoints with invalid IDs.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability allows remote attackers to crash the vLLM engine, leading to service disruption. It may also enable unauthorized memory writes, potentially causing data corruption or further security issues depending on the environment.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability primarily impacts system availability and integrity by allowing denial-of-service or memory corruption. While not directly violating GDPR or HIPAA, it could lead to service disruptions or data exposure if exploited, potentially affecting compliance with availability and security requirements under these regulations.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM to version 0.29.0 or later to address the improper input validation issue. Implement input validation for stop_token_ids in API requests. Restrict access to vLLM endpoints to trusted users only. Monitor for unusual engine crashes or logit buffer errors.

Useful 0 Not Useful 0