vLLM EngineCore Denial of Service via Unchecked min_tokens

Executive Summary

This is a denial-of-service vulnerability in vLLM versions before 0.31.0. When an attacker sends a request to the /inference/v1/generate endpoint with a null max_tokens and an extremely large min_tokens value, the EngineCore input thread crashes. The /health endpoint still returns 200, but all subsequent generation requests hang until a manual restart.

Useful 0 Not Useful 0

Detection Guidance

Monitor the /inference/v1/generate endpoint for requests with null max_tokens and unusually large min_tokens values. Check if the /health endpoint returns 200 while generation requests hang. Inspect vLLM logs for uncaught VLLMValidationError exceptions or EngineCore input thread crashes.

Useful 0 Not Useful 0

Impact Analysis

Unauthenticated attackers can exploit this to make the vLLM service unresponsive, blocking all generation requests. The service appears healthy via /health, but no new requests are processed until manually restarted. This disrupts AI inference workloads and requires downtime to recover.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability primarily impacts availability by causing a denial of service, which could lead to service disruptions. For GDPR, this may affect the right to access or availability of services. For HIPAA, it could disrupt access to protected health information systems. However, the vulnerability does not directly expose data, so confidentiality requirements remain unaffected.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM to version 0.31.0 or later to apply the fix. Implement input validation to ensure min_tokens <= max_tokens even when max_tokens is null. Restrict access to the /inference/v1/generate endpoint to trusted clients only.

Useful 0 Not Useful 0