ag2-multimodal-input
Pass
Audited by Gen Agent Trust Hub on Sep 23, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted external content, including images, audio, video, and documents (PDFs, Markdown, etc.). This content is processed by AI agents, creating a potential surface for indirect prompt injection where instructions embedded within the media (e.g., text in an image or metadata in a PDF) could attempt to influence the agent's behavior.
- Ingestion points: Media inputs are ingested via
ImageInput,AudioInput,VideoInput, andDocumentInputfactory methods inSKILL.md. - Boundary markers: None are explicitly documented in the instruction set, relying on the underlying agent's default safety boundaries.
- Capability inventory: The skill enables network operations (fetching from URLs), local file reading (fetching from paths), and provider-specific API interactions for file management.
- Sanitization: The provided examples do not include explicit content sanitization, validation, or filtering of the processed media files.
- [EXTERNAL_DOWNLOADS]: The skill includes examples for fetching media from external sources such as YouTube and generic image URLs. These are documented as standard functionality for the intended multimodal processing use case.
Audit Metadata