The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted media content (video, audio, or transcripts) provided via the source parameter. This content is a vector for adversarial instructions designed to influence the model's extraction behavior or bias its results.
Ingestion points: Untrusted media is ingested via the source input (SKILL.md).
Boundary markers: The instructions lack explicit boundary markers or safety guidelines to differentiate between the agent's instructions and potentially malicious content within the media files.
Capability inventory: The skill utilizes the Gemini API for agentic video understanding, which involves parsing the provided source based on a goal.
Sanitization: No specific sanitization or filtering logic is defined for the input content beyond basic output normalization.
[DATA_EXFILTRATION]: The source input allows for local media paths to be provided to the agent. This creates a potential surface for data exposure where an attacker could attempt to point the agent to sensitive system files (e.g., configuration or credential files) under the guise of media analysis, potentially resulting in the contents being transmitted to the external Gemini API.
[EXTERNAL_DOWNLOADS]: The skill references and interacts with the Gemini API and Google AI Studio (ai.studio). These are recognized well-known services and the interaction is consistent with the skill's stated purpose.

Agentic video understanding