paidf-augmentation
Pass
Audited by Gen Agent Trust Hub on Sep 1, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media files (images and videos) and text descriptions from various sources including local paths, S3, GCS, Azure, and HTTP. These inputs are analyzed by a Vision-Language Model (VLM), and the resulting descriptions are used by an LLM to generate generative prompts. This processing chain represents an indirect prompt injection surface where malicious instructions embedded in input media or its metadata could influence the behavior of the captioning or generation stages.\n
- Ingestion points: Untrusted media is ingested via the
data[*].inputs.rgbfield defined inSKILL.mdandreferences/configuration-schema.md.\n - Boundary markers: The prompt templates provided in
references/captioning-strategy-guide.mdlack explicit delimiters or instructions to ignore embedded commands within the VLM-generated descriptions.\n - Capability inventory: The pipeline has the capability to perform network requests to remote model endpoints and write files to both local and cloud storage (e.g.,
output.video,output.caption,output.metadata) as described inreferences/pipeline-operations.md.\n - Sanitization: There is no evidence of input sanitization for media metadata or VLM-generated descriptions before they are interpolated into LLM prompts.
Audit Metadata