SGLang Unbounded Memory Allocation via File Upload

Executive Summary

CVE-2026-109225 is a memory exhaustion vulnerability in SGLang through version 0.5.21. It allows unauthenticated remote attackers to crash the system by uploading oversized files to the /v1/audio/transcriptions endpoint. The files are fully read into memory before validation, causing denial of service.

Useful 0 Not Useful 0

Detection Guidance

Monitor for unusually high memory usage on systems running SGLang v0.5.21 or earlier, particularly on endpoints handling /v1/audio/transcriptions. Check for large multipart uploads or concurrent requests exceeding normal traffic patterns. Use system monitoring tools like top, htop, or ps to track memory consumption by the SGLang process.

Useful 0 Not Useful 0

Impact Analysis

This vulnerability can lead to service disruption by consuming excessive memory, crashing the HTTP/tokenizer worker, and denying access to the transcription endpoint. Concurrent attacks may degrade overall system performance and availability.

Useful 0 Not Useful 0

Compliance Impact

This vulnerability primarily causes denial-of-service by exhausting memory, which could disrupt services handling sensitive data. GDPR requires ensuring availability of processing systems, so prolonged outages may violate Article 32. HIPAA mandates safeguards against unauthorized disclosure or disruptions, which this flaw could exacerbate if exploited against healthcare systems.

Useful 0 Not Useful 0

Mitigation Strategies

Upgrade SGLang to a patched version that enforces upload size limits on the /v1/audio/transcriptions endpoint. Implement reverse proxy or application-level middleware to enforce Content-Length checks and maximum body size limits. Block or restrict access to the vulnerable endpoint if not in use.

Useful 0 Not Useful 0