The Agent Skills Directory

[COMMAND_EXECUTION]: The skill involves executing the cinefoundry CLI tool and interpolates a user-controlled <id> into the shell command output path (projects/<id>/breakdown). If the agent does not properly sanitize this identifier, it could be exploited for shell command injection.- [INDIRECT_PROMPT_INJECTION]: The skill is designed to analyze external video content, which provides an ingestion surface for indirect prompt injection. Malicious instructions could be embedded within on-screen text or visual elements in the video to influence the agent's behavior during the breakdown process.
Ingestion points: Video files (reference.mp4) uploaded by the user or fetched from external links.
Boundary markers: The instructions include some behavioral constraints (e.g., "do not let the model guess brand names it cannot read"), but these are not secure boundaries against adversarial content.
Capability inventory: Execution of the cinefoundry CLI and interaction with vision model endpoints.
Sanitization: No explicit sanitization or validation of the project ID or the data extracted from the video frames is specified.- [EXTERNAL_DOWNLOADS]: The skill's description suggests it can be used when a user "links a reference ad," which implies that the agent will perform network operations to download video files from external and potentially untrusted sources.

ad-reference-breakdown