byted-tos-image-process
Pass
Audited by Gen Agent Trust Hub on Sep 21, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The
scripts/image_understanding.pyscript provides a surface for indirect prompt injection by processing external images from Volcengine TOS through a Vision Language Model (VLM). Instructions hidden within the visual content of an image could potentially influence the agent's behavior during analysis. - Ingestion points:
scripts/image_understanding.pyingests untrusted image data from external TOS buckets for processing by thedoubao-seed-1.6-visionmodel. - Boundary markers: The script does not implement explicit boundary markers or instructions to the model to disregard instructions found within the source image.
- Capability inventory: The skill possesses significant capabilities to write files both locally and back to TOS storage via scripts such as
scripts/image_process.py,scripts/image_watermark.py, andscripts/image_draw.py. - Sanitization: The output from the image understanding model is processed and displayed without specific sanitization or filtering for malicious instructions.
Audit Metadata