Denial of Service via Memory Exhaustion in vLLM

Executive Summary

This vulnerability affects vLLM versions up to 0.17.0. It allows remote attackers to cause a Denial of Service by exhausting server memory. The issue occurs in the AsyncMediaIO.fetch_audio and AsyncMediaIO.fetch_image functions in multimodal/inputs.py, which fetch media URLs using aiohttp without enforcing a maximum response size. Attackers can exploit this by providing URLs to very large files.

Useful 0 Not Useful 0

Detection Guidance

To detect this vulnerability, monitor for unusually high memory usage on systems running vLLM up to 0.17.0. Check for processes consuming excessive memory when handling media URLs. Use commands like 'ps aux --sort=-%mem' to identify high-memory processes or 'top -o %MEM' for real-time monitoring. Inspect network traffic for large or continuous data streams from untrusted sources.

Useful 0 Not Useful 0

Impact Analysis

If you use vLLM up to 0.17.0, an attacker could send requests with URLs to large files, causing your server to run out of memory. This would make your service unresponsive or crash, leading to downtime and potential loss of availability for legitimate users.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability primarily impacts availability by causing denial-of-service through memory exhaustion. It does not directly lead to data disclosure or unauthorized access, so its impact on compliance like GDPR or HIPAA is likely minimal unless service disruption affects regulated data processing.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade vLLM to a version beyond 0.17.0 where the issue is resolved. Implement size limits on media file downloads in the AsyncMediaIO.fetch_audio and AsyncMediaIO.fetch_image functions to prevent memory exhaustion.

Useful 0 Not Useful 0