The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as video files, audio files, and marketing briefs, which could contain hidden malicious instructions intended to subvert the agent's behavior.
Ingestion points: Data is ingested from Google Cloud Storage URIs and local files for processing by multimodal Gemini and Imagen models in references/examples.md and references/core-capabilities.md.
Boundary markers: The provided instructions and code examples lack boundary markers or explicit delimiters to separate user instructions from untrusted external data.
Capability inventory: The skill is configured with powerful tools including Bash, Write, and Edit (as defined in SKILL.md), which could be abused if the model is successfully subverted by an injection.
Sanitization: There is no evidence of input validation or sanitization for the content processed by the Vertex AI models before it is used in generation or analysis tasks.

vertex-ai-media-master