CVE Alert: CVE-2026-109225 – sgl-project

SGLang through 0.5.21 contains an unbounded memory allocation vulnerability that allows unauthenticated remote attackers to exhaust memory by uploading oversized files to /v1/audio/transcriptions. Attackers can send concurrent large multipart uploads, which are fully read into memory before validation, to crash the HTTP/tokenizer worker and deny service.

**Risk verdict:** This is a serious availability risk for exposed inference services; KEV status, SSVC exploitation state and EPSS are not provided, so confirmed exploitation urgency cannot be assessed.

**Why this matters:** Successful attacks could interrupt transcription and other workloads sharing the affected serving process, causing service outages, queue backlogs and recovery costs. The likely goal is disruption rather than data theft or modification.

**Most likely attack path:** An attacker can reach the transcription endpoint over the network without credentials or user interaction, then submit oversized uploads—potentially in parallel—to consume worker memory. The impact is confined to the vulnerable service’s availability; the supplied metrics do not indicate direct cross-boundary impact or inherent lateral movement.

**Who is most exposed:** Internet-facing or broadly accessible AI inference deployments that enable audio transcription are most at risk, especially shared services with limited resource isolation.

Alert on unusually large or concurrent requests to the transcription endpoint.

Monitor worker memory, restarts, out-of-memory events and sudden drops in request throughput.

Correlate upload volume and failed requests with service degradation.

Mitigation and prioritisation

Upgrade to a maintainer-confirmed fixed release; the supplied record does not identify one.

Until patched, restrict endpoint access and apply strict upload-size and concurrency limits at the ingress.

Disable transcription if it is not operationally required; test controls against legitimate audio workloads.

Check KEV, SSVC and EPSS data before final prioritisation; reassess if exploitation indicators emerge.

A considerable amount of time and effort goes into maintaining this website, creating backend automation and creating new features and content for you to make actionable intelligence decisions. Everyone that supports the site helps enable new functionality.

If you like the site, please support us on Patreon or Buy Me A Coffee using the buttons below.