The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user data, including creative briefs, visual references, and narrative descriptions, which serves as an entry point for indirect prompt injection attacks.
Ingestion points: Creative briefs, user-owned images, brand rules, and reference examples provided in Phase 1 and Phase 2 (SKILL.md).
Boundary markers: No specific delimiters or safety warnings are instructed for the agent to use when interpreting these external inputs.
Capability inventory: The skill uses sandbase_run for external API interactions and "local media tools" for file operations like concatenation, mixing, and frame extraction (SKILL.md, Phase 7).
Sanitization: While the skill mandates using schema-defined arguments for API calls, it lacks specific instructions for sanitizing user inputs before they are processed by local media utilities or interpolated into generation prompts.
[COMMAND_EXECUTION]: The skill instructs the agent to utilize "local media tools" for deterministic operations such as frame extraction, concatenation, and audio mixing. This typically involves the execution of command-line utilities (e.g., ffmpeg), creating a surface for command injection if parameters derived from user-supplied filenames or metadata are not properly sanitized.
[EXTERNAL_DOWNLOADS]: The skill interacts with external generative models provided by SandBase. While this is the primary purpose of the skill, it involves the transmission of user-provided references to external infrastructure. The instructions include positive security constraints, such as treating references as private data and preventing the unauthorized cloning of real faces or voices.

video-generator