together-fine-tuning

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.70). The vlm_finetune.py workflow (url_to_base64 and the --sample-image-url option) explicitly downloads arbitrary image URLs (and references/data-formats.md shows the same URL-to-base64 pattern), then embeds that untrusted external content into training and inference datasets, so third-party content from the open web can be ingested and materially influence the fine-tuned model's behavior.