SGLang API Denial of Service via Client-Controlled Request ID

Executive Summary

SGLang through 0.5.22 has a denial of service vulnerability where API clients can abort other users' ongoing generation requests by sending a short client-controlled request ID (rid). Attackers exploit this by submitting streaming requests with single hex character rids and disconnecting, which triggers prefix-matching logic that terminates all other clients' requests.

Useful 0 Not Useful 0

Detection Guidance

To detect this vulnerability, monitor for API requests with unusually short request IDs (rids) like single hex characters (e.g., 'a', '1', 'f') being sent to streaming /generate endpoints. Check logs for disconnect events triggering abort requests that use prefix matching logic. Inspect scheduler and tokenizer manager logs for unintended request cancellations.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability allows attackers to disrupt legitimate users' ongoing tasks by prematurely terminating their requests. This can cause service unavailability, wasted computational resources, and degraded user experience for all users relying on the system.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability primarily causes denial of service by allowing attackers to abort other users' requests through prefix-matching logic. While it does not directly expose or leak data, it could disrupt service availability for multiple users simultaneously. This may impact compliance with regulations like GDPR (availability requirements under Article 32) and HIPAA (security rule integrity and availability requirements) by potentially causing service disruptions that affect data processing or access.

Useful 0 Not Useful 0

Mitigation Strategies

Update SGLang to a version beyond 0.5.22 to address the denial of service vulnerability. Implement rate limiting and request validation to prevent abuse of the rid parameter.

Useful 0 Not Useful 0