Executive brief
ms-swift is a framework for fine-tuning large language models and multimodal AI systems. The vulnerability allows unauthenticated attackers to force the server to make requests to arbitrary URLs (including internal services and cloud metadata endpoints) by supplying malicious image, audio, or video URLs. This could expose sensitive configuration data, enable lateral movement within networks, or compromise cloud credentials.
Technical details
The vulnerability is a server-side request forgery (SSRF) in ms-swift 4.5.2's OpenAI-compatible API endpoint that processes multimodal media. The vulnerable component (swift/template/vision_utils.py) fetches content from image_url, audio_url, or video_url parameters without validating or filtering the destination URLs or following redirects safely. Unauthenticated attackers can exploit this over the network by crafting requests with arbitrary URLs pointing to internal services, cloud metadata endpoints (e.g., AWS metadata), or other restricted resources. An attacker can retrieve sensitive data from internal systems or use the server as a proxy for further attacks. A patch is available in newer versions of ms-swift.
Affected products
- ModelScope ms-swift 4.5.2
Timeline
- 2026-09-04: disclosed